Your agent can read a contract in French, reconcile a ledger and explain a stack trace before lunch. On Monday afternoon your finance lead corrects it once: refunds above £75 need approval, not above £100. On Tuesday morning it approves a £90 refund without asking anyone.
The model is working as designed. It brings general intelligence, learned slowly from the world, to whatever is in front of it, and by Tuesday the correction is no longer in front of it. Multiply that by every correction your team makes in a week, and your agent spends its days repeating mistakes your people have already fixed.
What it lacks is a second kind of intelligence, one that learns fast from your work and acts on it while the work happens. Duplex is that intelligence. With it beside your model, Monday’s correction applies on Tuesday, and on every day after.
Your AI brings general intelligence to every session.
Your AI brings general intelligence.
General intelligence is a remarkable thing to have on tap. A frontier model learned language, knowledge and reasoning from the world’s writing, and it brings all of it to questions it has never seen: a contract clause, a customer in a hurry, a bug no one has met before. No system your team builds will match that breadth, and none needs to.
That breadth was learned slowly, in pretraining, before the model was deployed. Once it is in use, it stays as it was trained, however much of your work passes through it. It knows what the world wrote down, and nothing about what your team decided this morning. A better version arrives when its maker ships one, trained on the world again rather than on your work.
Teams fill the gap with context. A prompt adds what matters for one request, and retrieved passages add what looks relevant. Both help, and both are gone when the request ends, so a correction has to be put back in front of the model every time it matters, by someone who knows to do it. What is missing is something that learns at the speed of the work.
Your AI learns general knowledge, language and reasoning in pretraining, before it is deployed.
Duplex adds a second intelligence.
Duplex learns from each result your team confirms, as it happens, and acts on it from the next request, in every session and for every agent you allow. Your model’s weights stay as they are, and Duplex needs no GPU: it runs on ordinary servers, so the accelerators stay with your AI, for the new work only it can do.
The two divide the work by what each does best. What Duplex has learned, it answers itself: in About 97 ms, with no model call, against About 2,231 ms when your AI works it out. What is new, or only looks familiar, goes to your AI, briefed with what Duplex knows about the case. The result comes back, and Duplex learns from that too.*
Briefed means your AI starts from what Duplex knows: the rules that apply, how your team handles cases like this, the correction from Monday. It spends its general intelligence on what is new, instead of working out again what your team decided weeks ago. And the longer the two work together, the more of the routine work Duplex takes on, so your AI’s calls go where they count.
The usual way to teach a model something is a retraining project: gather the corrections, train, test, release. The correction your team made today waits for every step of it. Duplex learns it as it happens, and what it learns works beside your AI rather than inside it, so your team can inspect it, correct it, remove it or carry it to another model.
Your team confirms a correction on a live request, while the work is happening.
The brain pairs a slow learner with a fast one.
The brain works with two systems: a slow one for general knowledge, and a fast one for each new experience. It is why you can learn a new colleague’s name in a moment without unlearning anything else you know. Each system protects the other: the fast one learns without disturbing what took years to build, and the slow one stays steady while the fast one keeps up with the day.
Your AI is general intelligence, learned slowly, like the neocortex. Duplex is runtime intelligence: it learns fast from your work, then decides and acts on it, like the hippocampus. Together they are two intelligences in one system. Your model stays as dependable as the day it was deployed, and Duplex learns as fast as your work changes.
The analogy stops at learning. Duplex also decides whether to answer, hand over or decline, plans and acts within your rules, and carries the work between sessions, agents and models. It is an intelligence of its own, beside your AI, and it gives your AI what it cannot do alone.
The difference shows when the work changes. On a stream of 3,000 decisions in which the rules changed five times, the model on its own was right on 2.3% of the decisions those changes affected. The two together were right on 92.5%, and over the whole stream they made 90.3% fewer model calls.*
Keep your AI for what only general intelligence can do: the new case, the hard question, the request no one has seen. Let Duplex take what your team has already settled, and learn from every result along the way. One brings what the world knows, the other what your work has taught it, and the system they make together gets better every day it runs.