jobseekers

Rehearsal: prove your AI agent works before it ships. An AI agent can tell a customer "your refund is issued" while the records say otherwise. Most agent evaluations ask another model whether the transcript looks right, and a confident agent can talk its way to a pass. Mocks agree with whatever the agent does, and production is too risky to practise on. Rehearsal is continuous integration for AI agents. It runs your agent on a working copy of your real application, with fictional data, and decides every outcome by reading the application's own database. No model grades the result. How it works: 1. World Compiler. Point Rehearsal at a repository and its container images. The compiler reads the source, starts the application in an isolated sandbox, finds its API, seeds fictional data, chooses the tools each role may use, and writes jobs. It then proves every job can be passed and can be failed before the job counts. The result is a versioned, resettable practice world. 2. Rehearsals. Your agent works through real jobs while a simulated customer changes their mind mid-task and requests fail on purpose, for example a write that times out after it already went through. 3. Database-verified outcomes. Read-only checks on the application's records decide each outcome: goal reached, no duplicate side effects, and claims that match the data. "Refund issued" counts only if the refund is there, exactly once. 4. Readiness verdict. "Ready to pilot" or "not yet", with every measure linked to the episodes behind it, and failed episodes explained in plain words. 5. Fix my agent. A loop that writes the next agent version from failures. It keeps a version only if it verifies more episodes without breaking a rule, then proves the best one on held-out jobs. 6. Agent Factory. No agent yet? Rehearsal drafts one from the world's tools and your policies, hardens it, proves it on held-out jobs, and packages it as a standalone agent.py with a PROOF.md of every score. 7. Pull request checks. "rehearsal gate" and a GitHub Action rehearse a changed agent next to the version on main. The check fails on a broken rule or any regression. It ships as one package, rehearsal-kit, for Python (pip install rehearsal-kit) and TypeScript (npm install -g rehearsal-kit). The package contains the rehearsal CLI, an SDK, and an MCP server with 20 tools for AI assistants. A web console and a docs site take a new user from install to their first verdict. A real build: we pointed Rehearsal at Vikunja, an open-source task manager. The World Compiler ran 10 build steps, with 880 Muse model calls and $46.77 of model spend recorded by role. It published a world with 4 validated scenarios, 8 scoped agent tools and 3 roles. Each scenario stress-tests something real: the app refusing or failing, a customer pushing for something forbidden, a request missing a detail, and a write that times out after it succeeded.