An instrument, not a platform.

It does not replace your stack.

Every fix lands as a real GitHub pull request labelled agent-etna. Your branch protection, CODEOWNERS, required reviewers and CI run on it the same way they run on every other PR. There is no proprietary runtime, no rewrite, and nothing to migrate to. Your model API and your repository stay where they are.

The conventional position is that agent tooling has to own the runtime to be useful. We think the opposite. The work an experienced team has already put into review and CI is the most valuable part of the system — the tool's job is to feed it, not to bypass it.

Numbers, not vibes.

Every change ships with a pass-rate, a regression count and side-by-side outputs against the previous version. The numbers are derived, not invented:

• Baselines are computed from your system prompt, not handwritten.
• Prompt variants run in the shadow against live traffic, so drift is detected before users see it.
• Comparisons across model versions are reported per-test, not in aggregate.
• Cost and latency are tracked at the percentile, not the average — averages hide the failures.

A signature on every change.

Every approved fix carries a cryptographic signature, verifiable on a public page. If someone asks three years from now whether a particular change came from your team, the answer is in the audit log and not in a Slack thread. Built on an open, public trust standard — no proprietary trust root — so the verification does not depend on us still existing.

Everything the UI does is an API call.

POST /api/sandbox/apply opens a PR. GET /api/sandbox/pr/:n reads its checks. POST /api/fix/start kicks the auto-fix pipeline. The MCP server exposes the same surface to other agents. If you would rather drive Agent Etna from your own code than from the dashboard, that path is first-class — it is the same path the dashboard takes.