How repeated work
becomes a program.
Follow one request through the loop. Click Next to continue—no typing or setup needed.
Explore the programs yourself (optional)
The walkthrough above needs only clicks. Here you can also edit the sample and choose a program directly.
Try it on editable code
Everything below runs in your browser. No code, query, or API key is sent to a server.
Loading example…
Program input
Buttons select programs explicitly. The demo executes deterministic code; it does not pretend a trained router chose the program.
Program output
What we have measured.
| Measure | Usual agent | With program |
|---|---|---|
| Model calls | 8 | 5 |
| Reported tokens | 33,954 | 30,963 |
| Repository tests | Passed | Passed |
| Bug-specific check | Passed | Passed |
The original code failed the same bug-specific check. Both patches fixed it. The context program triggered twice. This was a single run per approach on an issue we had already seen; repeatability and general savings remain unproven.
Other findings and limits
An earlier, different program that automatically ran test and diff checks used 63 calls versus 65 across nine complete comparisons, but 7.9% more tokens. That result does not establish a cost benefit.
In the fixed held-out search replay, six searches supported a bounded context proposal. Two were followed by a supported source-read command: the program’s window fully covered one read and partially covered the other. Most historical shell commands were outside this first parser’s supported subset.
The expanded-read program showed no covered follow-up turns in its held-out sample. The patch-review packet was compatible with three Git-only turns, but was not an exact command match.
These are range-coverage and compatibility results, not demonstrated saved calls. Extra context has a token cost. Live paired tests are still needed to establish task quality and net savings. Methods and evidence ↗