Max — trading systems engineer, Moscow
Your agents ship faster than your team can check them.
I build the layer that decides which agent output is safe to merge — evals, traces, tool limits, and the few places a human still has to look.
The gap you already feel
Almost every engineer on your team uses coding agents daily. Almost none of them trust the output enough to put it in production without reading it line by line — and nobody has time to read it line by line.
So the work quietly splits in two: output that got checked, and output that got approved. Both look identical in the pull request. You find out which was which during an incident.
This is not a model problem. It's a missing harness — the scaffolding around the model that assembles context, restricts tools, records what happened, and tests the result. Most failed agent rollouts die here, not on model quality.
What I build
- evals
Tests for agent behaviour, not just code
A suite that runs on every prompt or workflow change and fails loudly when the agent starts doing something new.
- traces
A record of every run you can actually read
Tool calls, injected context, decisions. So a bad merge takes twenty minutes to explain instead of a day.
- limits
Explicit boundaries on what agents may touch
Allowlisted tools, default deny, no autonomous writes to anything that costs money or wakes someone up.
- gates
The short list of things a human must approve
Small enough that people actually do it, instead of rubber-stamping everything at 4 minutes per review.
Why me
I run Mercury , an algorithmic trading system I built and operate solo. In trading, unvalidated output isn't a code-quality opinion — it's money leaving the account within the hour. Evals, traces and hard limits aren't process there, they're the reason the thing still exists.
I'm bringing that discipline to teams whose agents write more code than anyone can read.
How we can work
- 2 weeks
Harness audit
I take your real agent runs, find where verification is missing, and hand back a working eval suite plus tracing on one critical workflow. fixed fee — scoped on first look
- 2–3 mo
Build the layer
Full validation path for agentic changes: evals in CI, trace retention, tool policy, and the review gates. Your team owns it when I leave. monthly retainer — scoped on audit
Start with one failed run
Send me a single agent run that went wrong — the trace, the diff, whatever you have. I'll tell you whether it was a prompt problem, a missing tool permission, or stale context, and what would have caught it. No charge, no call required.
[email protected]