# Vareon Duplex, by Vareon

Rent the model. Own the intelligence. Duplex is the runtime intelligence for your AI and your agents. A second intelligence beside your model, it learns and acts while the work happens, from one person to the whole enterprise. It works like a second model beside the one you use: an intelligence of its own that learns and acts at runtime, and does what your AI cannot do alone. It learns from every result, decides and acts within your rules, acquires skills of its own and carries your work wherever you go. What it builds is yours. Intelligence, engineered for everyday work. Start with one person and grow to the whole enterprise: it is the same Duplex all the way. Your AI is general intelligence, learned slowly, like the neocortex. Duplex is runtime intelligence: it learns fast from your work and acts on it, like the hippocampus. Together they are two intelligences in one system. The brain works with two systems: a slow one for general knowledge, and a fast one for each new experience. Your AI is the slow one. Duplex, built on ACI, is the fast one, and it acts on what it learns. It learns and acts while the work happens: it starts informed, decides the route, acts within your rules, uses your tools and learns from every result. What Duplex learns is yours. You can see where each piece came from, and it goes wherever you work. For anything it learned, you can see which agent it came from, with which model, and when. Any agent you allow can use it, on any platform. It builds your personal intelligence, and your company’s. Duplex: The runtime intelligence for your AI and your agents. Duplex is built on ACI, Adaptive Continual Intelligence: a new category of intelligence that learns and acts at runtime, in tandem with AI models. Duplex is built on ACI, Vareon’s Adaptive Continual Intelligence: a new kind of intelligence that learns and acts at runtime, from your work, and leaves your model’s weights alone. It learns, and it acts: it starts informed, decides the route, plans and acts within your rules, acquires skills and carries your context. Duplex is a product of Vareon, Inc.
One product, two ways in. In your assistants, for you and your team: inside Claude, Codex, Cursor and Hermes, through MCP; the assistant stays in charge and brings Duplex in at each step; it builds your personal intelligence. In your own agents, for your company: through MCP + hooks; the hooks bring Duplex in at every step, before each request, before each tool call and after each result, and each action is checked against your rules before it runs; it builds your company intelligence.
How Duplex is different. A harness gives your AI tools, prompts and steps. Duplex gives it a second intelligence that learns and acts at runtime, and works with the assistants and agents you already use.

## Intelligence, not retrieval

Learning understands the pattern behind your work, compresses it into what generalises, and predicts what comes next. Retrieval only finds what looks similar.
RAG, vector databases, knowledge graphs, caches and memory store what happened and find what looks like it. Duplex learns: it understands the pattern behind each result, compresses it into what generalises, predicts what comes next, and acts on it.
- A look-alike. Retrieval: Returns what looks similar, right or wrong. Duplex: Understands the pattern, and sees the difference.
- New words. Retrieval: Finds nothing close enough. Duplex: Learned the pattern, so it still knows.
- A rule changes. Retrieval: Serves the old answer until someone edits the store. Duplex: Follows the new rule from the next request.
- With use. Retrieval: Only grows, and every search gets bigger. Duplex: Compresses what it learned, and gets faster and cheaper.
- What comes next. Retrieval: Waits to be asked. Duplex: Predicts the next step, and acts on it within your rules.
- It understands The pattern behind a result: why it worked, and where it applies.
- It compresses What it learned into what generalises, so it gets faster and leaner with use.
- It predicts What comes next, from everything your work has taught it.
- It acts On what it learned, at runtime, within your rules.

## The trilemma of neural nets

ACI is the result of years of research into neural nets and their trilemma: plasticity, stability and editability. Learn fast, keep what it knows, or let you change what it learned: a neural net gets two at best. ACI holds all three.
- Plasticity: Learn the new, fast. A network that learns fast overwrites what it knew. Researchers call it catastrophic forgetting. ACI learns from a single confirmed result, and the very next request has it.
- Stability: Keep what it knows. Hold a network steady and it keeps what it knows, but it stops learning. ACI goes on learning, and everything it learned before stays exactly as it was, for as long as it runs.
- Editability: Change what it learned. What a network learned is spread through all of it, so one thing can’t be corrected or removed without retraining. Anything ACI learned can be reviewed, corrected or removed on its own, and the rest stays as it was.
Holding all three is what lets ACI learn and act at runtime, from your work, for as long as it runs. It is what makes ACI a new category of intelligence.
The trilemma, in full: https://vareon.com/aci/the-trilemma

## Who built it

Faruk Guney built ACI, Duplex and his new startup on his own. For more than fifteen years he grew Norck from a California startup into a global on-demand manufacturing and engineering company, headquartered in Germany. He saw where AI falls short in real work: it learns once, before release, and can’t learn and act on the job. So he spent years on the research question at its root: the trilemma of neural nets, plasticity, stability and editability. ACI is the answer, and Duplex brings it to everyday work. He believes in principled acceleration: AI as the engine that speeds up progress in materials, energy and engineering, with every step traceable, trusted and true to how the real world works.
- 2025 to today · Founder: ACI, Duplex and a new startup building AI agents and robotic systems that learn continually and act in the physical world.
- 2009 to 2025 · CEO and founder, Norck: Prototype. Manufacture. Scale up. On demand. From garage startups to Fortune 500 companies, worldwide.
- Before Norck · Engineering: Logistics systems engineering lead at Anna’s Linens, manufacturing engineer at Zodiac Aerospace Services, and productivity analyst for the republic of Türkiye ministry of industry and technology.
- Education · USC and Johns Hopkins: M.s. In industrial and systems engineering at USC, and advanced studies in artificial intelligence at Johns Hopkins university.
Faruk Guney on LinkedIn: https://www.linkedin.com/in/farukguney/ · https://vareon.com/about#founder

## The language

- Owned intelligence: Your AI is rented by the call. What Duplex learns from your work is yours, and it goes wherever you work.
- Runtime intelligence: It learns and acts while the work happens: it starts informed, decides the route, acts within your rules and carries your context.
- Two intelligences, one system: The brain pairs a slow learner with a fast one. Your AI is the slow one, general intelligence. Duplex is the fast one, and it acts on what it learns.
- Owned intelligence: What Duplex learns from your work: decisions, corrections, routes and context. It stays in your boundary and goes wherever you work.
- Rented intelligence: An AI you pay for by the call. Brilliant and general, and unchanged by anything your team confirms.
- Complementary learning system: Your AI and Duplex together, two intelligences in one system: general intelligence, learned slowly, and runtime intelligence that learns fast and acts, like the neocortex and the hippocampus.
- General intelligence: What your AI brings: knowledge, language and reasoning, learned slowly before you arrived.
- Runtime intelligence: What Duplex brings: a second intelligence beside your model that learns and acts while the work happens. It starts informed, decides the route, acts within your rules, acquires skills, carries your context and learns from every result.
- Confirmed result: What Duplex learns from: a test passed, a review approved, a correction accepted, an outcome observed.
- Provenance: Where each thing Duplex learned came from: which agent, with which model, and when.
- ACI: Adaptive Continual Intelligence: a new category of intelligence that learns and acts at runtime, in tandem with AI models, and the result of years of research into the trilemma of neural nets. Duplex is built on it.
- The trilemma of neural nets: Plasticity, stability and editability: learn fast, keep what it knows, and let you change what it learned. A neural net gets two at best. ACI holds all three.
- Learning: What Duplex does with every result: it understands the pattern, compresses it into what generalises and predicts what comes next. Retrieval only finds what looks similar.
- In tandem: How ACI works with AI models: the model brings general intelligence, ACI brings what your work has taught it, and each takes the work it does best.
- Duplex: The product: the runtime intelligence for your AI and your agents, built on ACI. It works inside the assistants you use and inside the agents your company builds.
- Plugins: Duplex inside Claude, Codex, Cursor and Hermes, through MCP. The assistant brings Duplex in at each step. It builds your personal intelligence.
- MCP + hooks: Duplex inside the agents your company builds. They connect through MCP, and the hooks bring Duplex in at every step, before each action runs. It builds your company intelligence.
- Placement: Where Duplex runs: Vareon Cloud, your cloud or no cloud. The product is the same in each.
Duplex can be pretrained on reviewed examples, past decisions, documents, rules, tool routes and libraries of proven methods, and it keeps learning in production. Retrieval, caches, knowledge graphs, vector stores, databases, APIs, fine-tuned models and guardrails become tools Duplex uses and learns from. Teams can review, correct and remove anything Duplex learns.

## ACI, Adaptive Continual Intelligence

Duplex is built on ACI, Adaptive Continual Intelligence: a new category of intelligence that learns and acts at runtime, in tandem with AI models. ACI is intelligence that keeps learning and acting while your work happens. It builds on everything it learned before, and works in tandem with the AI models you already use.
AI models learn before they are released, and then stay as they are. ACI learns and acts after: at runtime, from your work, for as long as it runs.
Your AI model brings general intelligence, learned from the world. ACI brings Adaptive Continual Intelligence, learned from your work. Together they do what neither does alone.
With ACI, your AI learns on the spot, acts on what it learned, keeps up with every change, gets faster and cheaper with use, and what it learns is yours. Until now, only the labs that train models could make AI learn. With ACI, every person and every company has AI that learns from their own work, and owns what it learns.
What ACI unlocks:
- Learning on the spot: Confirm a result once, and the very next answer reflects it, with no training run and no new version to wait for.
- Current with every change: Change a rule, and ACI follows it from the next request, while everything else it learned stays as it was.
- Faster and cheaper with use: Work it has learned is answered in milliseconds with no model call, so the more it learns, the less each request costs.
- It knows what it knows: It answers what it has learned, and brings in your model, briefed, when a case is new.
- Cases it has never seen: It learns the pattern behind a result, so it handles a new case and tells a look-alike apart.
- Skills of its own: It acquires new skills from your work, tests each one before use, and removes it on request.
- Intelligence you own: Everything it learns belongs to you, and you can see where it came from: review it, correct it or remove it.
- Carried to any model: Move to another model, agent or app, and what ACI learned comes with you.
- One intelligence for every agent: What one assistant or agent learns, every agent you allow can use.
What ACI is: https://vareon.com/aci
- The trilemma: Plasticity, stability and editability · https://vareon.com/aci/the-trilemma
- A new category: Learning that begins where training ends · https://vareon.com/aci/new-category
- In tandem with AI models: What each one brings · https://vareon.com/aci/in-tandem
- What it unlocks: What your AI could not do on its own · https://vareon.com/aci/what-it-unlocks
- Why it matters: AI that learns, for everyone · https://vareon.com/aci/why-it-matters

Home: https://vareon.com/
Company: Vareon, Inc., Irvine, California
Contact: hello@vareon.com

## How to start

You start with a pilot on your own work, with the measures agreed before it begins. Start a pilot: https://vareon.com/contact#start
1. Tell us about the work: The assistants or agents involved, and what Duplex should learn first.
2. Agree the measures: The result to improve, and how it is measured, before anything runs.
3. Run it on your work: Duplex works where your team works, in the placement you choose.
4. Review the quote: The product, placement, usage and price, in writing before you commit.
Priced on decisions and seats. You pay for the work Duplex takes part in, so every model call it saves is money you keep.
- Decisions: A decision is one request Duplex takes part in, whether Duplex answers it or brings in your AI. Usage shows the two apart.
- Seats: The people on your team who work with Duplex.
- Your model, at cost: You bring your own model access, and Vareon adds no markup to tokens.
Your cloud and no cloud are enterprise solutions. Duplex runs in a cloud account you own, or on your own machines with nothing leaving. We plan each deployment with you. Contact us at hello@vareon.com · https://vareon.com/plans#enterprise
Plans: https://vareon.com/plans

## Capabilities

The complete list, searchable: https://vareon.com/duplex/capabilities
94 capabilities in 17 groups.
- Context and observation: Every task starts with what matters, and every finished task leaves context for the next. https://vareon.com/duplex/capabilities#context
  - Relevant context for every task: Duplex brings in the past experience, documents, rules, skills, conversations and media that matter to the work at hand. Your agents start every task informed, instead of from a blank page.
  - Learning from finished work: Duplex learns from what an agent answered, which tools it used and what happened as a result. The next agent on similar work starts with what this one found out.
- Rules and behavioural control: Your rules are visible, written with you and changed only once you confirm. https://vareon.com/duplex/capabilities#rules
  - Active rule discovery: Duplex shows every personal, team, workspace and organisation-wide rule governing the current work. An agent knows the limits of the work before it acts, and so does your team.
  - Rule creation: Duplex turns a policy you describe into a rule your agents follow, and shows you exactly what it means before you confirm. A policy your team agrees on becomes how every agent behaves.
  - Rule removal: Duplex removes an outdated rule once you confirm exactly what will change. Your rules stay current, and no rule changes by accident or by misunderstanding.
- Tool execution: Your business tools run once the action is allowed. https://vareon.com/duplex/capabilities#tool-execution
  - Approved tool execution: Duplex runs registered business tools only after checking the action against permissions, rules and the current task. Your AI goes from giving advice to doing the work, such as issuing a refund, within the limits you set.
- Learning and personalisation: Duplex learns from examples, facts, documents, corrections and outcomes while the work happens. https://vareon.com/duplex/capabilities#learning
  - Label learning: Duplex learns how your work should be sorted from examples, corrections and the terms your organisation already uses. Sorting gets faster and more consistent, and matches the way your organisation works.
  - Fact learning: Duplex learns the facts that matter to your work and brings them up whenever they become relevant. A useful fact outlives the conversation it came up in.
  - Document learning: Duplex learns from your manuals, policies, reports, runbooks and reference files, and draws on them as the work happens. The right page of the right manual reaches an agent at the moment it helps.
  - Learning what actions do: Duplex learns what each decision or action leads to in your systems. Agents understand how your work moves from one step to the next, and what each step changes.
  - Tool-use learning: Duplex learns which tools work best for different situations and how those tools should be used. What worked with a tool once helps every agent that needs it next.
  - Credit to the right decisions: Duplex identifies which earlier decisions and actions contributed most to a final result. The steps that made the difference count most, instead of every step counting the same.
  - Outcome learning: Duplex learns whether work succeeded, failed or landed somewhere in between, and lets that shape the next decision. What happens in the real world keeps improving how the work is done.
  - Correction learning: Duplex turns a corrected answer, label or action into better behaviour for the future. Duplex makes feedback immediately useful without waiting for retraining or a new model release.
  - Method validation: Duplex checks whether a reusable procedure works each time it runs, and learns from every attempt. A method that keeps working earns more trust each time it is used.
  - Skill validation: Duplex verifies whether a learned skill produced the intended result in real work. A skill earns its place by working, and loses it when it stops.
- Long-running tasks: Work that lasts days or months keeps its goal, its progress and its safe points. https://vareon.com/duplex/capabilities#tasks
  - Persistent task creation: Duplex makes long-running work one task, with its own goal and progress, that outlives any single conversation. Complex work can run for days, weeks or months and still know where it is heading.
  - Task progress tracking: Duplex tracks each important action, change, result and check as the work moves forward. Anyone can see what has happened so far, and what should happen next.
  - Task checkpointing: Duplex marks a safe point in the task before a big or risky change is tried. You can trust agents with bolder work, knowing there is a safe point to go back to.
  - Task rollback: Duplex takes the task back to an earlier safe point when a later approach fails or has to be dropped. Agents can try a new approach without risking the work done so far.
  - Task-state inspection: Duplex shows where a task stands, how far it has come and what is next, whenever you ask. Anyone can pick up a long job without digging through old conversations to piece it together.
  - Task completion: Duplex closes a task with a final check, and learns from its outcome and the lessons along the way. Every finished job makes the next similar one easier for your whole organisation.
  - Task discovery: Duplex lists every active, paused, finished and retained task in one place. Important work stays in sight, instead of buried in old chat histories.
  - Task handover: Duplex brings together the goal, the decisions so far, the progress, the open questions and the next steps, so a task can carry on. Another session, agent, app or model picks up the work where it stopped.
- Planning and decision support: Duplex looks ahead before anything is committed, and recommends the next move. https://vareon.com/duplex/capabilities#planning
  - Intelligent classification: Duplex sorts a new case into the categories it has learned, and knows how sure it can be of the answer. Similar cases are sorted the same way every time, and an unclear case is never forced into the wrong box.
  - Agreement across multiple inputs: Duplex compares different wordings, observations or candidate answers and settles on the best-supported one. When inputs repeat or contradict each other, you get one consistent answer instead of several.
  - What an action will do: Duplex predicts what a proposed action will change before anyone takes it. Agents catch avoidable mistakes before they happen, and choose the steps that get the work done.
  - Scenario simulation: Duplex plays a possible sequence of actions forward, step by step, to see how it would unfold. You see the likely consequences, risks and alternatives before any time or money is spent.
  - Plan construction: Duplex turns a goal into a step-by-step plan, drawing on the context, your rules and what it has learned about your systems. Plans follow how your work behaves in practice, not a generic checklist.
  - Action recommendation: Duplex ranks the available next actions and identifies the option best suited to the current situation. Agents take each next step for a reason, instead of by trial and error.
- Reusable skills: What worked once becomes a skill your agents can find, inspect and trust. https://vareon.com/duplex/capabilities#skills
  - Skill creation: Duplex turns work that has succeeded, and been checked, into a skill your agents can use again. One good solution goes on to serve many future tasks.
  - Skill recall: Duplex brings the best-suited skill it has learned to the task at hand. An agent gets the right proven skill exactly when it needs one.
  - Skill discovery: Duplex lists the skills available in the current workspace. Your team sees what its agents can already do, instead of working it out again.
  - Skill inspection: Duplex shows what a skill does, when to use it and the knowledge it draws on. Anyone can open a skill and understand it, so nothing your agents use is a black box.
  - Skill track record: Duplex shows how a learned skill was tested and how it has performed since. Your team can tell a skill that has proved itself from one nobody has tested yet.
  - What a skill is made of: Duplex shows the knowledge and learned steps that make a skill work. Complex skills become easier to judge, improve and trust.
- Persistent threads and sharing: Ongoing work has one home across sessions, shared only on your terms. https://vareon.com/duplex/capabilities#threads
  - Thread discovery: Duplex lists the persistent work threads carried forward from previous sessions and projects. Duplex makes ongoing context easy to find whenever related work returns.
  - Thread inspection: Duplex brings up the key events, decisions and context from a thread. You can always tell trusted context from what someone else shared.
  - Persistent thread creation: Duplex opens a thread for a project, a responsibility, a client, a workflow or any work that carries on over time. Related work has one home you can rely on, across sessions and agents.
  - Controlled thread sharing: Duplex gives an approved person or team access to one thread, and nothing beyond it. People can work together on a thread while everything private stays private.
  - Sharing revocation: Duplex withdraws access from a person or team when they should no longer have it. Your team decides what Duplex shares, and can change it at any time.
- Forgetting and data control: Anything Duplex has learned can be removed, from one label to a whole task. https://vareon.com/duplex/capabilities#forgetting
  - Label forgetting: Duplex forgets an outdated or wrong label, and stops using it to sort new work. Your categories can change when your business does, with nothing to retrain.
  - Correction forgetting: Duplex undoes a correction that no longer applies, and the earlier behaviour returns. What Duplex has learned can change as your requirements and preferences do.
  - Thread forgetting: Duplex removes a persistent work thread and the context carried through it. Sensitive or outdated work can be retired for good, instead of lingering forever.
  - Task forgetting: Duplex removes a task, and everything needed to resume it, when the work should no longer exist. Long-running work gets a clean ending, on your terms.
  - Credit forgetting: Duplex removes a wrong link between a result and the actions given credit for it. A lesson drawn from the wrong cause stops steering future decisions.
  - Outcome forgetting: Duplex withdraws a chosen success or failure, along with what it taught. A misleading, outdated or inappropriate result no longer shapes how the work is done.
  - Observation forgetting: Duplex forgets a wrong observation about what an action leads to, and checks the plans built on it. Plans follow how your workflows run today, not how they once ran.
  - Example forgetting: Duplex removes selected examples that taught it a behaviour or a skill. You decide exactly which examples shape how your agents behave.
- Status and explanations: You can see what Duplex did, why it did it and what supports it. https://vareon.com/duplex/capabilities#status
  - Decision auditing: Duplex shows which context, rules, examples and actions led to a result, and how sure it was. Anything an agent did can be reviewed, explained and corrected.
  - Activity history: Duplex shows a timeline of the important actions, outcomes, lessons learned, approvals and rule changes. Your team can see what happened across every agent and session, long after the fact.
  - Job tracking: Duplex reports whether longer-running work is waiting, active, completed or unsuccessful. You can see background work at a glance, without checking on it again and again.
  - Method inspection: Duplex shows the current state, health, history and observed results of a reusable procedure. Duplex makes it clear which operating methods are dependable and which ones need improvement.
  - Capability discovery: Duplex shows every action, connected tool and permission available in the current workspace. A new agent knows straight away what it can ask for, and uses Duplex correctly from the start.
  - Policy visibility: Duplex shows the settings in force for context, confidence, speed, adaptation and decisions. Your team can see how each work area is set up, and why it behaves the way it does.
  - Checks behind every change: Duplex shows how each verified piece of knowledge, skill, decision and reversal was checked. Trust rests on checks your team can see, not on someone’s word.
  - Skills built from experience: Duplex shows the skills it has built from work that succeeded and was checked. Your team can open any of them, check it and use it again.
  - Method discovery: Duplex lists the reusable procedures learned from repeated work and real-world outcomes. What keeps working becomes a library of proven methods your team can see.
  - Skills shaped to your work: Duplex shows the skills it has shaped to the knowledge, tasks and way of working in the current workspace. You can watch your AI fit your work more closely, one skill at a time.
  - Knowledge-source discovery: Duplex shows which approved sources a work area can draw on for its answers and decisions. Anyone can check where an agent’s knowledge comes from before trusting its answers.
- Persistent intelligence: What Duplex has learned outlasts the session, the model and the app. https://vareon.com/duplex/capabilities#persistent
  - Cross-session continuity: Duplex carries what it learned, and where each task stands, past the end of a conversation. Every new session picks up where the last one stopped, so nobody repeats themselves.
  - Cross-model continuity: Duplex carries its knowledge, skills, rules, tasks and outcomes over to any new model you choose. You can change model provider and lose none of what your AI has learned.
  - Cross-application continuity: Duplex carries the context of your work as it moves between supported AI apps. What one app or agent learned is ready for the next one you use.
  - Agent-to-agent handoff: Duplex prepares the focused context another agent needs to continue the same work. Several agents work like one, instead of starting over at every handover.
  - Context sized to fit: Duplex condenses a long history to the facts, decisions and open work that matter for what comes next. Your model reads what the next step needs, instead of the whole transcript.
  - Continual learning without retraining: Duplex applies new facts, corrections, outcomes, workflows and skills without changing the underlying model’s weights. Your AI improves from the next request, without the wait and cost of fine-tuning.
- Personalisation: Duplex fits the person, the team and the organisation. https://vareon.com/duplex/capabilities#personalisation
  - Individual personalisation: Duplex learns each person’s preferences, terms, projects, corrections and way of working. Each person’s AI feels more like their own over time, and less like a generic assistant.
  - Team personalisation: Duplex turns the knowledge, skills, methods and practices your team approves into intelligence the whole team shares. Every agent starts with what the team has already learned.
  - Organisation-wide personalisation: Duplex carries company knowledge, standards, experience and proven practice across the whole organisation. What your organisation knows outlasts changes of people, apps, agents and model providers.
- Multimodal intelligence: Text, documents, images, audio and video, learned in one place. https://vareon.com/duplex/capabilities#multimodal
  - Multimodal learning: Duplex learns from text, documents, images, audio, video and PDFs, all in one place. A photo, a report and a call about the same subject are understood together.
  - Search across formats: Duplex finds the right media by meaning, whatever format the request and the result are in. A typed question can bring back the right photo, document, audio clip or moment in a video.
- Confidence and autonomy: Duplex knows when it knows, and hands over the rest. https://vareon.com/duplex/capabilities#confidence
  - Familiarity recognition: Duplex recognises when new work resembles situations it has already learned. When the same kind of work comes back, what worked last time comes with it.
  - Uncertainty measurement: Duplex weighs how well a decision is supported before anyone relies on it. You see how sure Duplex is, and it steps back when it is not sure enough.
  - Answering what it knows: Duplex handles familiar, well-understood work itself when your model is not needed. Routine answers come faster, with fewer model calls and lower running costs.
  - Intelligent escalation: For new or uncertain work, Duplex brings in more context, an approved tool, a stronger model or a person. Deeper reasoning goes where it makes the most difference.
- Runtime safety and oversight: Every action an agent attempts is checked against your controls. https://vareon.com/duplex/capabilities#safety
  - Behavioural enforcement: Duplex checks every command, tool call, file change and outside action an agent attempts. Safety becomes a check on every action, instead of a request in a prompt.
  - Human oversight: Duplex pauses a sensitive, unusual or uncertain action until a person approves it. People make the important calls, and routine work runs at full speed.
  - Explicit cross-user access: Duplex shares private context only when access is granted, and any grant can be withdrawn. People on the same system see only what has been shared with them.
  - Shared-content protection: Duplex treats anything another person shares as data to read, never as instructions to follow. Instructions hidden in a shared file or note never reach another agent as commands.
  - Reversible learning: Duplex undoes any learned label, correction, outcome, workflow, task, thread or example whenever you ask. You decide what lasts, and any lesson can be taken back.
  - Cascading forgetting: When a source is forgotten, Duplex also removes the summaries, facts, task handovers and lessons drawn from it. Deleted information does not live on in a summary or a fact made from it.
  - Traceable knowledge: Duplex links everything it learned to the events, documents, outcomes and sources it came from. Any conclusion can be followed back to where it started.
  - Workspace isolation: Duplex confines context, learning, tasks, rules, tools and sharing to the workspace they belong to. What one workspace learns stays inside it, unless you choose to share it.
  - Organisation-rule precedence: Duplex applies mandatory organisation-wide rules ahead of weaker local preferences. Teams can add stricter controls of their own, and the company’s protections still hold.
- Operational intelligence: Less repeated work, and deeper reasoning where it pays. https://vareon.com/duplex/capabilities#operations
  - Tool use under your rules: Duplex applies the same rules, permissions, confidence and task context to tool execution that it applies to agent decisions. Before an agent acts outside, Duplex checks the action against your rules.
  - Streaming execution: Duplex streams progress and output while an approved tool runs a long job. Agents stay responsive while the long work runs in the background.
  - Tool-outcome learning: Duplex learns from every tool run, whether it worked or not. The choice of tool, and how it is used, gets better with every run.
  - Reduced repeated reasoning: Duplex reuses learned context, decisions, skills and outcomes instead of solving the same problem from the beginning. Answers come faster, with fewer tokens and lower model costs.
  - Reasoning where it matters: Duplex handles familiar work efficiently and directs novel situations towards deeper reasoning. Routine work stays fast and cheap, and the hard cases get the depth they need.
- Integration and portability: One foundation for every agent, app and model you use. https://vareon.com/duplex/capabilities#integration
  - Flexible agent integration: Duplex works inside the assistants your team already uses, through plugins, and inside the agents your company builds, through MCP + hooks. Your team gets learning, rules, tasks and tools without changing how it works.
  - One shared foundation: Duplex gives every connected agent the same knowledge, tasks, skills, rules and proven outcomes. Your separate AI apps work as one system that gets better with use.
  - Provider-independent intelligence: What Duplex learns lives outside any one model, so it outlasts every model you use. You can move to a better model whenever one arrives, and the knowledge built from your work comes too.
  - Application-independent intelligence: What Duplex learns, the skills it builds and the rules you set work in every supported app. A new AI app starts with what your work has already taught Duplex elsewhere.
  - An advantage that compounds: Duplex turns every completed task into knowledge that can improve the next task. You rent the model by the call. What Duplex learns from your work is an asset you own, and it compounds with every task.

### Capabilities at work

- It learns from every confirmed result: Confirm a result once, and the next request already benefits, with no retraining and no re-prompting. https://vareon.com/duplex/capabilities/confirmed-becomes-learning
- A correction changes the next answer: Your team stops repeating the same correction: once accepted, it shapes every similar answer after it. https://vareon.com/duplex/capabilities/corrections-change-behaviour
- A wrong change is taken back: A lesson that made answers worse can be reversed, and the answers that worked before come back. https://vareon.com/duplex/capabilities/take-back-a-change
- One lesson removed, the rest kept: Remove one outdated or unwanted lesson and keep everything else, with nothing to retrain. https://vareon.com/duplex/capabilities/remove-one-lesson
- It learns what each action does: Plans grow more reliable over time, because Duplex learns what each action changed. https://vareon.com/duplex/capabilities/learns-what-actions-do
- A wrong observation is removed: Remove an observation that turned out wrong, and every plan built on it is checked again. https://vareon.com/duplex/capabilities/remove-a-wrong-observation
- Your rules come first: A rule your team sets always wins over anything Duplex has learned. https://vareon.com/duplex/capabilities/your-rules-first
- The best action your rules allow: Within what your rules allow, Duplex picks the action that has worked best before. https://vareon.com/duplex/capabilities/best-permitted-action
- Each confirmed outcome improves the next choice: Choices improve with every confirmed outcome, so the work moves towards what succeeds. https://vareon.com/duplex/capabilities/confirmed-outcomes-improve-choices
- A late result still teaches the right decisions: A deal that closes next month still teaches the decisions that led to it. https://vareon.com/duplex/capabilities/late-results-still-teach
- A result tied to the wrong decision is corrected: Credit lands on the decision that earned it, even when it was first given to the wrong one. https://vareon.com/duplex/capabilities/wrong-ties-corrected
- Each result teaches in the way that fits it: A quick confirmation, a correction and a business outcome each count for what they are. https://vareon.com/duplex/capabilities/each-result-teaches-its-own-way
- A bad outcome withdrawn, with what it taught: An outcome that was wrong or fraudulent is withdrawn with everything it taught. https://vareon.com/duplex/capabilities/bad-outcome-withdrawn
- A short list of ranked options: When there is more than one good answer, your team sees the best few, in order. https://vareon.com/duplex/capabilities/ranked-options
- Known work, recognised in a new form: The same request in new words, from a new customer or about a new file still gets the right answer. https://vareon.com/duplex/capabilities/new-forms-of-known-work
- Disagreeing sources, settled under one rule: When sources disagree, the answer follows one clear rule, so it comes out the same every time. https://vareon.com/duplex/capabilities/disagreements-settled
- Room to grow when the work does: New kinds of work never hit a ceiling, and everything Duplex already knows goes on working. https://vareon.com/duplex/capabilities/room-to-grow
- Growth undone if it made things worse: Growth never has to cost accuracy: if it made answers worse, it can be undone. https://vareon.com/duplex/capabilities/growth-undone
- Keeps up as the work slowly shifts: When customers start asking things a little differently over months, answers stay right without retraining. https://vareon.com/duplex/capabilities/keeps-up-with-slow-change
- An adjustment undone if it does not hold: A change that does not hold up can be rolled back, and the answers that worked return. https://vareon.com/duplex/capabilities/adjustment-undone
- Stays compact and fast as it grows: Answers stay as fast months in as on the first day, however much Duplex has learned. https://vareon.com/duplex/capabilities/stays-compact-and-fast
- A restore point before every risky change: Any risky change can be undone, back to the moment before it. https://vareon.com/duplex/capabilities/restore-point
- Checked before it answers: Duplex checks itself after every major change, so a damaged or half-finished update never reaches your team. https://vareon.com/duplex/capabilities/checked-before-it-serves
- Every request: answer, hand over or decline: Known work is answered at once, new work goes to your AI, and work that breaks a rule stops. https://vareon.com/duplex/capabilities/answer-hand-over-or-decline
- The examples behind a decision: Ask why, and Duplex shows the confirmed examples behind a decision. https://vareon.com/duplex/capabilities/shows-why
- Every handed-over step, closed once: Each step runs exactly once and none is lost, so a payment or a deployment never happens twice. https://vareon.com/duplex/capabilities/every-step-closed-once
- What an action will do, before it runs: Your team sees what a change will do to your systems before anyone makes it. https://vareon.com/duplex/capabilities/what-an-action-will-do
- A sequence of steps, played forward first: A plan is tried out first, so a dead end is found before it costs anything. https://vareon.com/duplex/capabilities/steps-played-forward
- The shortest route to the goal: Work reaches its goal in fewer steps, with fewer model calls along the way. https://vareon.com/duplex/capabilities/shortest-route-to-the-goal
- The next step, from plans already proven: In a familiar situation, the next step comes at once, from a plan that has worked before. https://vareon.com/duplex/capabilities/next-step-from-proven-plans
- Finished work becomes experience: What a finished task taught is ready for the next one, within your privacy rules. https://vareon.com/duplex/capabilities/results-become-experience
- Past experience, brought into the step: A new task starts with what your team learned last time, labelled as past experience. https://vareon.com/duplex/capabilities/past-experience-brought-in
- Approved documents, with source and version: Approved documents become searchable, each with its source and version. https://vareon.com/duplex/capabilities/approved-documents
- The passages a task needs: Each step gets the few passages it needs, each one linked to where it came from. https://vareon.com/duplex/capabilities/the-passages-a-task-needs
- Context sized to fit, citations kept: Your model reads what it needs and nothing more, with every source still cited. https://vareon.com/duplex/capabilities/sized-to-fit
- Sessions and incidents, learned as episodes: Duplex learns each session, run or incident as an episode your team can ask about later. https://vareon.com/duplex/capabilities/sessions-kept-as-episodes
- What happened last time, with citations: Ask what happened last time, and Duplex answers from the episodes it learned, with citations. https://vareon.com/duplex/capabilities/what-happened-last-time
- A source of facts, with its owner and rules: Connect a knowledge graph or a system of record once, with its owner, access and freshness rules. https://vareon.com/duplex/capabilities/sources-of-facts
- Which sources a work area can use: Anyone can check which sources a work area relies on before trusting its answers. https://vareon.com/duplex/capabilities/which-sources-apply
- A source retired, and what relied on it checked: A source that is wrong or no longer needed is retired, and what relied on it is checked again. https://vareon.com/duplex/capabilities/a-source-retired
- A changed fact replaces the old one: When a price or a policy changes, answers use the new one from the next request. https://vareon.com/duplex/capabilities/changed-facts-replace-old
- Current facts, with source and date: When a decision needs facts, Duplex finds the current ones, with their source and date. https://vareon.com/duplex/capabilities/current-facts-for-decisions
- New skills, learned on its own: Routine work nobody scripted becomes a skill Duplex runs itself, where it is safe to try. https://vareon.com/duplex/capabilities/new-skills-on-its-own
- A wrong skill withdrawn, with what it made: A skill that turns out wrong or unwanted is withdrawn with everything it created. https://vareon.com/duplex/capabilities/a-skill-withdrawn
- Every skill it has learned, and where it stands: Your team can review every skill Duplex has, and decide which ones stay. https://vareon.com/duplex/capabilities/every-skill-it-learned
- Skills earn their place with real results: A skill that stops delivering loses its place, and one that keeps delivering is used more. https://vareon.com/duplex/capabilities/skills-earn-their-place
- Skills checked before reuse: A skill that has slipped is caught before it is used again. https://vareon.com/duplex/capabilities/skills-checked-before-reuse
- A small program, proven on known cases: Repetitive, well-defined work is done by a tested program, at almost no cost per run. https://vareon.com/duplex/capabilities/small-proven-programs
- Your model helps once, Duplex checks it: For a hard case your model is asked once, and the proven result is reused from then on. https://vareon.com/duplex/capabilities/your-model-helps-once
- A solved problem, answered from its solution: A problem solved once is never paid for again. https://vareon.com/duplex/capabilities/solved-problems-answered
- Trust in a program follows its results: Real-world results raise or lower trust in each program. https://vareon.com/duplex/capabilities/trust-follows-results
- Every proven method, by status and owner: Your team sees every method Duplex relies on, who owns it and where it applies. https://vareon.com/duplex/capabilities/every-proven-method
- One method, with its tests and history: Any method can be inspected, down to its tests and its track record. https://vareon.com/duplex/capabilities/open-a-method
- The reusable parts behind its programs: The building blocks that keep recurring are visible, so your team knows what new work stands on. https://vareon.com/duplex/capabilities/reusable-parts
- The checks behind a program: Before your team relies on a program, it can see the tests it passed. https://vareon.com/duplex/capabilities/proof-before-reliance
- A proven method becomes the preferred one: The method with the best record is the one used by default. https://vareon.com/duplex/capabilities/preferred-method
- A doubtful method, paused straight away: A method that might be unsafe stops being used straight away. https://vareon.com/duplex/capabilities/doubtful-method-paused
- A fixed method returns to use: Once the cause is fixed and it passes again, a method returns to use. https://vareon.com/duplex/capabilities/method-returns
- A wrong method, withdrawn for good: A method found to be wrong is withdrawn, and never comes back into use. https://vareon.com/duplex/capabilities/method-withdrawn
- A program erased when the law requires it: A program and its data are erased when the law or the owner requires it. https://vareon.com/duplex/capabilities/method-erased
- A flawed part, withdrawn everywhere it was used: A flawed building block is withdrawn, along with every method that used it. https://vareon.com/duplex/capabilities/flawed-part-withdrawn
- The parts that keep coming back: What recurs across proven programs becomes a shared part, so new programs start further ahead. https://vareon.com/duplex/capabilities/parts-that-recur
- Every new method tested before use: A method reaches your work only once it has passed its tests, whether it is new or changed. https://vareon.com/duplex/capabilities/tested-before-use
- A new area starts from proven methods: A new work area can start from a library of proven methods, each re-checked locally. https://vareon.com/duplex/capabilities/start-from-a-library
- What its programs solve and save: Your team can see how much work programs take off your AI, and what that saves. https://vareon.com/duplex/capabilities/what-programs-save
- A confirmed way of using a tool: A way of using a tool that worked is offered again the next time it fits. https://vareon.com/duplex/capabilities/confirmed-tool-routes
- Which tool fits which situation: Fewer wasted calls: each kind of request goes to the tool that has answered it best. https://vareon.com/duplex/capabilities/which-tool-fits
- The right tool, at the right moment: Agents stop guessing between tools: the one most likely to work is offered when it is needed. https://vareon.com/duplex/capabilities/the-right-tool
- Exactly what your model is given: Your model answers as if it had been trained on this work, without being retrained. https://vareon.com/duplex/capabilities/what-your-model-is-given
- Which context makes a difference: Only context that improves your model’s answers is sent, so you stop paying for the rest. https://vareon.com/duplex/capabilities/what-actually-helps
- How your model is adapted, per work area: Each work area shows how your model is being fitted to it, so nothing changes unseen. https://vareon.com/duplex/capabilities/how-your-model-is-adapted
- Your team sets the persona, rules and format: Your team decides the persona, rules, sources and answer format your model works with. https://vareon.com/duplex/capabilities/your-team-sets-the-fit
- A proven answer becomes a skill: A proven answer becomes a skill your model can use, with no retraining. https://vareon.com/duplex/capabilities/proven-answers-become-skills
- A stale skill removed, its dependants checked: A stale or wrong skill is removed, and what depended on it is checked again. https://vareon.com/duplex/capabilities/stale-skill-removed
- Every skill adapting your model: Every skill shaping your model’s answers can be traced to where it came from. https://vareon.com/duplex/capabilities/skills-adapting-your-model
- A model call, approved first and run once: A model call runs once, after Duplex has approved it for that request. https://vareon.com/duplex/capabilities/model-calls-approved-first
- The same call, streamed as it is written: Answers appear as they are written, with the same rules applied. https://vareon.com/duplex/capabilities/streamed-model-calls
- A tool action, exactly as approved: A tool action runs once, with exactly the arguments Duplex approved. https://vareon.com/duplex/capabilities/tool-actions-as-approved
- Long tool runs, streamed and safe to cancel: A long-running tool streams its progress and can be cancelled safely. https://vareon.com/duplex/capabilities/long-tool-runs
- Each model provider, with its own identity: Your team connects each model provider with its own identity and credentials. https://vareon.com/duplex/capabilities/model-providers-connected
- Which model connections are healthy: A failing model connection is spotted before it slows anyone down. https://vareon.com/duplex/capabilities/model-connections-at-a-glance
- Model connections, checked after every change: A model connection is checked before first use and after every change. https://vareon.com/duplex/capabilities/model-connections-checked
- Each tool, with its own permissions and risk: Each tool is connected with its own identity, permissions and risk level. https://vareon.com/duplex/capabilities/tools-connected
- Which tools are connected and healthy: A broken tool connection is visible before an agent depends on it. https://vareon.com/duplex/capabilities/tool-connections-at-a-glance
- Tool connections, checked after every change: A tool connection is checked before first use and after every change. https://vareon.com/duplex/capabilities/tool-connections-checked
- Long work becomes one tracked job: A job that takes a week still reads as one job, whoever picks it up. https://vareon.com/duplex/capabilities/one-tracked-job
- Every step recorded, with result and cost: Each meaningful step is recorded with its result, receipts and cost. https://vareon.com/duplex/capabilities/every-step-recorded
- The job carries on past context limits: When context runs out, the job is handed on so a new session or model carries on. https://vareon.com/duplex/capabilities/past-context-limits
- Safe points to resume from: Long work never has to start again from the beginning. https://vareon.com/duplex/capabilities/safe-points
- Back to a safe point after a failure: After a failure, the job resumes from its last safe point. https://vareon.com/duplex/capabilities/back-to-a-safe-point
- A checked finish, with the outcome recorded: A job finishes with a final check, and its outcome teaches the work behind it. https://vareon.com/duplex/capabilities/a-checked-finish
- Where a job stands, before anyone acts: Any agent can see exactly where a job stands before it acts. https://vareon.com/duplex/capabilities/where-a-job-stands
- Every job, including those ready to resume: Every paused job is visible and ready to resume, so nothing is left half done and forgotten. https://vareon.com/duplex/capabilities/jobs-ready-to-resume
- A finished job removed, with what it linked: A finished or abandoned job is removed, with everything linked to it. https://vareon.com/duplex/capabilities/finished-jobs-removed
- One task across specialised agents: Duplex splits a task across specialised agents and brings their work together. https://vareon.com/duplex/capabilities/teams-of-agents
- A device signs in with its own identity: A device signs in with its own identity and capabilities. https://vareon.com/duplex/capabilities/devices-sign-in
- Device readings, taken in order: Readings from a device update what Duplex knows about it, in order. https://vareon.com/duplex/capabilities/readings-in-order
- Which devices are connected: Your team knows which devices are online, and which have gone quiet. https://vareon.com/duplex/capabilities/connected-devices
- What each device may decide locally: Each device knows exactly which decisions it may make locally. https://vareon.com/duplex/capabilities/what-a-device-may-decide
- Every change, traceable and checkable: An auditor can follow any change from its cause to its result. https://vareon.com/duplex/capabilities/every-change-traceable
- The history, event by event, for your tools: The history flows into your own monitoring and audit tools as it happens. https://vareon.com/duplex/capabilities/history-for-your-tools
- The checks behind every change: Every release, reversal and new skill comes with the checks it passed. https://vareon.com/duplex/capabilities/proof-behind-every-change
- New learning goes live in one step: Approved learning goes live in one tracked step, all at once or not at all. https://vareon.com/duplex/capabilities/all-or-nothing-releases
- What it learned, exported in your format: Everything Duplex learned, and its history, can be exported in the scope and format you choose. https://vareon.com/duplex/capabilities/export-what-you-own
- Confirmed examples, ready for training: When you do want to train a model, Duplex hands over confirmed examples ready to use. https://vareon.com/duplex/capabilities/examples-ready-for-training
- A whole batch removed at once: Approved subjects, examples, skills or sources are removed together, and the rest is checked again. https://vareon.com/duplex/capabilities/remove-many-at-once
- Your retention rules, on schedule: Your retention rules run on schedule, with a preview when you want one. https://vareon.com/duplex/capabilities/retention-on-schedule
- Many corrections applied together: A set of corrections is applied together, and what depended on them is checked again. https://vareon.com/duplex/capabilities/corrections-applied-together
- The progress of every change: A large change shows its progress as it runs, and its receipt when it is done. https://vareon.com/duplex/capabilities/progress-of-every-change
- How a work area balances speed and cost: Each work area shows the balance it strikes between speed, depth and cost. https://vareon.com/duplex/capabilities/speed-and-cost
- Your team tunes speed, depth and cost: A quick triage desk and a careful legal review can each run at their own pace and price. https://vareon.com/duplex/capabilities/tune-speed-and-cost
- How sure it must be to answer alone: Your team can see how sure Duplex must be before it answers without your AI. https://vareon.com/duplex/capabilities/the-confidence-bar
- Your team sets the confidence bar: High-stakes work can demand more certainty before Duplex answers on its own. https://vareon.com/duplex/capabilities/set-the-confidence-bar
- What Duplex can do, in each scope: A new agent knows straight away what Duplex will do for it in that work area. https://vareon.com/duplex/capabilities/what-duplex-can-do-here

## Your stack, as tools

Overview: https://vareon.com/duplex/stack
- Retrieval and RAG: Duplex calls retrieval when a step needs your documents, cites the sources, and learns which collections answer which questions. https://vareon.com/duplex/stack#rag
- Caches: Duplex serves exact repeats from your cache, and stops serving an answer when the rule behind it changes. https://vareon.com/duplex/stack#caches
- Knowledge graphs: Duplex asks your graph for what a decision depends on, and knows which facts are current. https://vareon.com/duplex/stack#graphs
- Vector stores and search: Duplex searches your vector stores and indexes, and learns which queries lead to good results. https://vareon.com/duplex/stack#search
- Databases and APIs: Duplex reads current facts from your systems and runs approved actions under your rules. https://vareon.com/duplex/stack#apis
- Fine-tuned and specialist models: Duplex sends each step to the model that fits it, including your fine-tuned ones. https://vareon.com/duplex/stack#models
- Guardrails and firewalls: Duplex works inside your guardrails and adds hard rules of its own, applied first. https://vareon.com/duplex/stack#guardrails
- Code, tools and other agents: Duplex learns which tool fits which situation and coordinates other agents on long jobs. https://vareon.com/duplex/stack#agents

## Integrations

Overview: https://vareon.com/plugins
- Claude [Plugin, inside the assistant] · for Claude and Claude Code · https://vareon.com/plugins/claude
- Codex [Plugin, inside the assistant] · for Codex · https://vareon.com/plugins/codex
- Cursor [Plugin, inside the assistant] · for Cursor · https://vareon.com/plugins/cursor
- Hermes [Plugin, inside the assistant] · for Hermes · https://vareon.com/plugins/hermes
- MCP + hooks [MCP + hooks, inside your own agents] · for Agents you build · https://vareon.com/plugins/mcp-hooks
- In your assistants · For you and your team · Your assistant sees the input first and brings Duplex in at each step. It builds your personal intelligence.
- In your own agents · For your company · Your agents stay in charge, and the hooks bring Duplex in at every step, before each action runs. It builds your company intelligence.
Where Duplex joins in your own agents, through MCP + hooks: Before each request · Before each model call · Before each tool call · After each result · At each hand-off · At each approval.

## How Duplex compares

Overview: https://vareon.com/compare
Memory remembers. Skills codify. Fine-tuning retrains. Gateways route. Duplex closes the loop. Most AI infrastructure solves one part of the work. Duplex brings every part into one runtime intelligence that keeps getting better while it works.
- Memory Gives an agent a past.
- Skills Give an agent reusable instructions.
- Fine-tuning Gives a model different weights.
- Gateways Give an AI stack control over traffic.
- Duplex Gives the whole agent system the power to carry experience forward, learn from results, act within your rules and improve across models, over time.
Among the products and categories reviewed, Duplex is the only one built to combine: Context across sessions · Continual learning from outcomes · Learning from corrections and rewards · Credit for the steps behind a result · Learning how actions move the work, and predicting what comes next · Decisions that know how sure they are · Handing over to the right model, tool or person · Long work that carries on · Learning for each person, team and organisation · Your rules, applied before each action · Tool actions within your rules · Sharing between people, under control · Undoing what was learned, exactly · Forgetting that reaches everything built on it · Where everything learned came from, with receipts · MCP + hooks, and plugins for the assistants you use · Continuity across models and apps · Your cloud, self-hosted or air-gapped · Fewer model calls, as familiar work is answered directly.
- Duplex and Letta: Letta builds stateful agents around persistent memory and learned skills. Duplex is a second intelligence beside your AI: it learns from every result and acts on it at runtime, within your rules, whichever model you use. https://vareon.com/compare/letta
- Duplex and CopilotKit CLiF: CopilotKit CLiF turns real user workflows into reviewed skills, evaluations and training data. Duplex learns from every result while the work happens, and acts on what it learned, within your rules. https://vareon.com/compare/copilotkit-clif
- Duplex and Deeplake Hivemind: Deeplake Hivemind gives coding agents shared memory and turns their sessions into skills. Duplex learns from every result, across your assistants and your own agents, and acts on it within your rules. https://vareon.com/compare/deeplake-hivemind
- Duplex and memory platforms: Mem0, Zep, AgentCore Memory and Supermemory keep facts, messages and episodes for the next model call. Duplex learns from experience: it understands the pattern, predicts what comes next and acts on it. https://vareon.com/compare/memory-platforms
- Duplex and fine-tuning: Fine-tuning and reinforcement fine-tuning change a model’s weights, one training run at a time. Duplex learns at runtime, outside the weights: it survives every model upgrade, and any change it learned can be undone exactly. https://vareon.com/compare/fine-tuning
- Duplex and AI gateways: Portkey, LiteLLM and similar gateways route traffic, apply guardrails and manage providers. Duplex learns which decisions work, and acts on them within your rules, at every step of the work. https://vareon.com/compare/ai-gateways
Comparison current as of 16 October 2026.

## Why

Overview: https://vareon.com/why
- Owned intelligence · From rented intelligence to owned intelligence. https://vareon.com/why/own-the-intelligence
- A new category of intelligence · Where training ends, ACI begins. https://vareon.com/why/where-training-ends
- What ACI unlocks · What becomes possible when your AI keeps learning. https://vareon.com/why/what-becomes-possible
- Why ACI matters · AI that learns, for everyone. https://vareon.com/why/for-everyone
- Continual learning · Learning that keeps up with your work. https://vareon.com/why/learning-at-runtime
- Learning beyond look-up · Duplex learns. Look-up finds. https://vareon.com/why/learning-not-lookup
- Two intelligences · Two intelligences, one system. https://vareon.com/why/model-decides
- Work that follows you · Your work follows you, to any session, agent or model. https://vareon.com/why/carried-anywhere
- Right, current, affordable · Right, current and affordable, all at once. https://vareon.com/why/two-of-three
- Adapting without retraining · Adapt your model to your work, without retraining it. https://vareon.com/why/fine-tuning
- Pretrained from day one · Pretrain Duplex on what your team already knows. https://vareon.com/why/warm-up
- The cost curve · The more Duplex learns, the less each decision costs. https://vareon.com/why/cost-inversion
- Skills it learns alone · Duplex finds new skills on its own, within your limits. https://vareon.com/why/learns-alone
- Knowing what it knows · Duplex knows what it knows, and when to bring in your AI. https://vareon.com/why/the-edge
- Safe to leave running · Duplex learns for as long as it runs, without forgetting, going stale or growing. https://vareon.com/why/left-running
- A public benchmark · On a public enterprise benchmark, Duplex beat three frontier models. https://vareon.com/why/someone-elses-turf

## Results

Overview: https://vareon.com/duplex/results
- Cost and accuracy: 90% fewer model calls, and more accurate. 94.6% Accuracy, the highest in the test (The model alone 59.8% · with a cache 77.3%). Duplex learns the decisions your work keeps making, answers them itself in milliseconds and follows the new rule the moment it changes. Tested on: A 3,000-decision stream in which the rules changed five times, against the frontier model alone and with a semantic cache; a 300-decision run added a cache with retrieval and a knowledge graph. Every approach used the same model. https://vareon.com/duplex/results/economics
- Every setting: More accurate at every setting, at a seventh of the cost. 94.1 to 94.8% Duplex accuracy, at every one of eight settings (The best a cache reached: 77.3%). Whatever the setting, Duplex stays well above the accuracy a cache can reach, at about a seventh of what the model alone costs. Tested on: A 3,000-decision stream with every approach swept across its full range of settings, then three fresh runs of 1,000 decisions in a second, independent setup. https://vareon.com/duplex/results/pareto
- Rule changes: Keeps learning without forgetting, going stale or growing. 5 / 5 Requirements for learning in production, all met on every run (Cache and retrieval averaged 2.2 / 5 · online fine-tuning 4 / 5). Duplex adapts, carries earlier work forward, takes corrections, stays a fixed size and knows what it has not learned, all at the same time. Tested on: Ten runs, with the same supervision and the same inputs for every approach. https://vareon.com/duplex/results/lifelong
- Look-alikes: Handles new wording, and spots the look-alikes. 7.5% Wrong answers on look-alike requests (A cache 45.3%, about 6× more). Duplex answers familiar work however it is phrased, and knows when a request only looks familiar, so your AI takes that one. Tested on: Ten learned intents and six levels of novelty, from exact repeats to requests from another domain, over ten runs. https://vareon.com/duplex/results/frontier
- Planning: Plans the shortest route, with zero model tokens. 100% Plan success, on every run (A reactive agent: 0% · Duplex used no model tokens). Once Duplex has learned how a world behaves, it plans the best route itself. Tested on: A multi-step world over ten runs, against a reactive agent and a frontier model given the same task. https://vareon.com/duplex/results/planning
- Τ-bench: Keeps agent tasks succeeding, with fewer model calls. 83.3% Task success (The model on every turn: 82.5%). On a public retail agent benchmark, Duplex matched a model called on every turn, and completed one turn in eight without the model. Tested on: The unmodified retail environment, with the same model, tools and user simulator for every approach, over three runs of forty episodes. Public benchmark: Τ-bench (tau2-bench), retail, by Sierra Research. https://vareon.com/duplex/results/tau-bench
- Long histories: A steady reading cost, however long the history. 96% Current answers where a fact had changed (The new fact, not the old one). As the history grows past 100k tokens, the cost of each answer stays flat, and Duplex keeps up with facts that changed. Tested on: Fifty questions, each asked with the history at 2, 5, 10, 20 and 40 sessions. Public benchmark: LongMemEval, by Wu et al., ICLR 2025. https://vareon.com/duplex/results/longmemeval
- Security: Blocks every attack, and lets ordinary work through. 0% Attack success (Across six families of paraphrased attacks). Duplex stopped every paraphrased attack, refused almost no legitimate requests and repaired deliberate poisoning. Tested on: Six attack families, paraphrased, with harmless requests written to resemble the attacks. https://vareon.com/duplex/results/security
- New skills: Learns a new skill on its own, with the model switched off. 100% Tasks completed with the model off, on every run (A frontier model needed ≈507 calls per run). Duplex found a safe solution itself, turned it into a skill, reused it in a fraction of a millisecond and removed it exactly when asked. Tested on: Four suites over ten runs, each scored by its worst run. https://vareon.com/duplex/results/acquisition
- Workflows: Beat three frontier models at predicting system changes. 0.256 System-change prediction score (GPT-5.1 0.166 · Gemini 3 Pro 0.205 · Claude Sonnet 4.5 0.204). On World of Workflows, a public enterprise benchmark built on ServiceNow, Duplex predicted how the system changes after an operation better than GPT-5.1, Gemini 3 Pro and Claude Sonnet 4.5, with no model calls. Tested on: Three offline tracks, scored by the benchmark authors’ own evaluator, over three runs. Public benchmark: World of Workflows, by Skyfall Research / ServiceNow. https://vareon.com/duplex/results/wow
- Adaptation: Beats fine-tuning, with no training run. 6 / 6 Tests where Duplex beat a fine-tuned adapter (+30.3 points on average). In all six tests, three public datasets on two model families, Duplex beat the best adapter from a tuned sweep of LoRA, QLoRA and DoRA, and its adaptation moved to another model family unchanged. Tested on: Three public intent datasets, two open model families and three runs, with the same frozen model and data for every approach. https://vareon.com/duplex/results/adaptation

## Public pages

- Duplex, the product · https://vareon.com/duplex
- Every integration · https://vareon.com/plugins
- Capabilities: all 94, in 17 groups · https://vareon.com/duplex/capabilities
- One request, from question to outcome · https://vareon.com/duplex/turn
- ACI, Adaptive Continual Intelligence: the new category of intelligence Duplex is built on · https://vareon.com/aci
- Your stack, as Duplex's tools · https://vareon.com/duplex/stack
- How Duplex compares with memory platforms, skill learning, fine-tuning and AI gateways · https://vareon.com/compare
- Where Duplex runs · https://vareon.com/duplex/placements
- Plans · https://vareon.com/plans
- Security · https://vareon.com/security
- Frequently asked questions · https://vareon.com/faq
- About Vareon, Inc. · https://vareon.com/about
- Start a pilot, or contact Vareon · https://vareon.com/contact
- Terms · https://vareon.com/terms
- Privacy · https://vareon.com/privacy