Repo: github.com/nikolas-sapa/co-train · Pre-registration: osf.io/2hfzt

What it is

co-train makes AI agents cheaper to run by turning their successful runs into verified, reusable programs.

Agents re-solve the same work from scratch on every run and pay frontier-model prices each time. Most tooling attacks this by cutting tokens. Tokens are a bad proxy for cost: one published study cut tokens 38% and the bill went up 7%. co-train measures dollars per solved task directly.

The result

Test plan fixed in the public repo on 25 Sep, before the first run; mirrored to OSF on 7 Oct (osf.io/2hfzt). I could not move the goalposts.

In those pre-registered runs, agents that wrote a short program to call tools solved at least as many tasks as agents using shell commands, at 6-8x lower cost per solved task.

The samples are small and the limits are published next to the numbers. The best run was 14 paired tasks, one rollout each, and the 6x end is a floor.

The part I published against myself

My first big result was a +64-point win. It was my own instrument. I found it was an error, kept digging, and found 13 instrument defects in total. Then I published the negative result on the original hypothesis alongside the cost finding.

Agent numbers lie quietly. Catching that is the skill co-train depends on, and it is the same skill behind my open-source fixes and the Convex security advisory.

What I'm building now

Capture an agent's successful programs as sandboxed, verified skills it reuses on new tasks, and train small models to prefer that path. Starting with SQL, because the answers are checkable.

The SQL skill-reuse system is built, with 3,400+ passing offline tests.
The first live end-to-end trial failed safely, before reuse. The fix is in progress.
About 200 commits since I started on 20 September 2026.

Where it stands

No users and no revenue yet. The measurement exists. The product does not.

I reached out to 79 founders and CTOs running agents in production. The first reply came from Isotopes AI, who shared their own data: in 392 finance-analysis sessions that needed retries, recovering from failures used 38.6% of their credits. That is the repeated work co-train is built to cut.

Stack

Python, PyTorch, TRL, MLX, Qwen, verifiers, MCP, BIRD SQL, Docker, pytest. Built with Claude Code and Codex. I decide what gets built, design the experiments and check the results.