Every Friday afternoon, someone on your team builds the weekly refund report. They pull the figures from the same systems, check them in the same order and send them to the same people. Nobody would call it hard. It is a skill, learned once and repeated, and every week it takes a capable person away from better work. Every team has a dozen tasks like it: the month-end reconciliation, the accounts for a new starter, the same three checks before every release.
Your AI can help, but it starts from nothing every time. It reads the instructions, reasons its way through each step and is paid for every call, and next Friday it does it all again. It can do the task. It never learns it.
Duplex acquires skills of its own. It learns a task by doing it, trying only the steps you permit, even with your AI switched off, and nobody has to write those steps down. It tests each skill before using it, runs it in a fraction of a millisecond from then on, with no model call, and lets it go the moment you ask.
On its own, Duplex tries the steps you permit, with your AI switched off.
It learns with the model switched off.
People learn skills by doing them. A new hire tries the steps, gets a few wrong, finds the order that works and, after a few weeks, does the job without thinking about it. A model can’t learn that way, because your work doesn’t change it. It brings broad knowledge to every task and reasons its way to the steps afresh, call after call.
There are good ways to get a task done today, and each one has a price. A script or a workflow is dependable and cheap to run once someone has written it, and it does only what it was written to do. A frontier model working the task out in context is flexible and needs no setup, and it pays for that flexibility on every run. Trial and error learns by doing, as people do, and takes a great many attempts to get there.
Duplex learns by doing too, and your AI can stay switched off the whole time. It tries the steps you permit, inside limits you draw as tight as the work needs, and finds a way that reaches the goal. Learning by doing is only safe when the doing has limits, so anything unsafe is refused before it runs. Across four suites over ten runs, each scored by its worst run, Duplex completed every task with the model switched off, in 2.8× fewer attempts than trial and error. It refused 627 unsafe actions in each run along the way, and still learned the safe way through.
The difference shows over a run. A frontier model solving the same task in context made about 507 model calls per run, working the task out again each time. Duplex learned it with the model switched off and made no model calls at all. For a task that comes round every week, that is the difference between a model bill that never ends and none at all. Duplex learns the task once, and your AI is free for the work that needs reasoning.
A frontier model working the task out in context makes a model call at every attempt.
It tests a skill before using it, and removes it on request.
A system that learns on its own is only useful if you can trust what it learns. The worry is fair: a shortcut that worked once and then gets used everywhere, or a habit nobody asked for that can’t be taken back. In a trained model, what it learned is spread through all of it, so no single lesson can be lifted out, and fixing one means training again.
Duplex tests a new skill before it uses it. A skill it has found goes into use only once it has passed its checks, and until then the work carries on as before. A shortcut that happened to work once is not enough. What reaches your systems and your customers is a skill that has already shown it works.
Once it has passed, a skill costs almost nothing to use again. Duplex repeated a learned skill in 0.26 ms, with no model calls, where a model working in context pays to reason through every step whenever the task comes round. Like everything Duplex learns, the skill belongs to you, and any agent you allow can use it. Work like the weekly report stops being something your AI works out and becomes something Duplex does.
And you can take any skill back. Processes change, systems are retired, and a step that was fine last quarter may not be fine now. Ask for a skill to go, and Duplex removes it and leaves everything else it learned as it was. In the tests, every learned skill was removed on request and checked bit for bit, so nothing of it was left behind.
That is what it takes to let a system learn on its own: limits you set before it starts, checks before anything new is used, and the power to take any of it back. With all three, the work your team repeats every week can become skills that Duplex finds, tests and runs in a fraction of a millisecond. Your AI takes the problems that need reasoning, your people make the judgement calls, and your team gets its fridays back.
A new skill Duplex has found is held back until it has passed its checks.