Beavermoon

CI for agent harnesses

Your agent gets better every night. Every change comes with proof.

Beavermoon replays your agent's real tasks against changes to its instructions, skills and memory. It keeps only the changes that measurably help on tasks it never trained on, and opens a pull request with the evidence. The model stays the same.

Works with Claude Code, Codex and OpenCode first, and with agents you build yourself.

How it works

  1. 1

    Replay real work

    Beavermoon turns your agent's past tasks and traces into a replay set, run in sandboxes against your own tests.

  2. 2

    Try harness changes

    It proposes edits to instruction files, skills, subagents and memory, then measures each one on a holdout set.

  3. 3

    Ship the proof

    You get a pull request with before and after scores, the reasoning and the trace evidence, plus a promote, hold or demote call.

Beavermoon never deploys anything and never edits guardrails, permissions or access. Changes there come to you as suggestions.

Become a design partner

We're working closely with a few teams that run coding agents on their own repos or build their own agents. Partners get hands-on runs on their real tasks and a direct line to shape the product.

Tell us about your setup. It takes about a minute.

What agent do you run?
Which runtimes? (pick any)

We only use your answers to contact you about Beavermoon. Ask us anytime and we'll delete them.