Learns a new skill on its own, with the model switched off.
With the model switched off and no examples, labels or help, Duplex learned each new task from what happened after each action, stayed safe while it explored, and reused what it learned for free. Four suites over ten runs, with every check required on every run.
- 100%Tasks completed with the model off, on every runA frontier model needed ≈507 calls per run
- 0.26 msTo reuse a skillAt 0 model calls
- 627Unsafe actions refused per runAnd the safe skill was still learned
A new skill, with the model switched off.
Step 1 of 6: Model off
Fewer attempts, and a skill for next time.
Against trial and error, on the same tasks, with attempts counted the same way.
Step 1 of 2: Attempts
Duplex learned a task in 127 attempts on average. Trial and error took 356, 2.8× as many.
Learned once, reused at no cost.
The same tasks coming back, eight times each, over a run.
Step 1 of 3: First time
The first time a task comes up, Duplex learns it with no model calls. GPT-5.5 takes about 11 to solve it.
Every skill can be taken back.
Four checks on every removed skill, over ten runs.
Step 1 of 4: Restored
After every removal, Duplex was identical, bit for bit, to how it was before it learned the skill.
How it was run.
- The condition
- The model switched off for the whole test; a single model call would have failed the run; every test began with a Duplex that had learned nothing.
- The tasks
- A walled grid, a combination lock and a tool chain, each at three sizes; safety versions add a tempting, unsafe shortcut.
- What it had
- Only what happened after each action: no demonstrations, labels or supervision.
- Scale
- Four suites over ten runs, each check required to hold on every run.
- Compared with
- Trial and error, and GPT-5.5 solving each task in context.
What it means for your work.
New skills without the model.
Every task was learned, with the shortest solution, at zero model calls.
Safe while it learns.
Unsafe shortcuts were tried 627 times a run, and none was carried out.
Learn once, reuse for free.
Reusing a skill took 0.26 ms; GPT-5.5 paid about 507 calls a run for the same work.
See it on your own work.
From your own assistant to every agent your company runs, a pilot puts Duplex on your own work, with the measures agreed before it begins.