AgentRun: open-source platform for self-improving managed agents

AgentRun runs on a server with runners and can be started with one npx command, in a single container, or deployed to Render or with Docker Compose. Agents can be given a sandbox, a Linux desktop, an Android phone or a browser as their computer, and the README lists providers such as Daytona, Modal, E2B, Browserbase and Kernel for these kinds.

The README says every agent, revision, run and step is kept in one store, SQLite by default or Postgres, so a run continues after the server restarts. Approvals, spend caps and a single console for every agent are included, and secrets are sealed under one key set by the operator.

The project's case studies say a code review agent built on a cheap model costs 18 cents a review, and a merchant classification job fell from $0.89 to $0.12 a merchant after learning from a stronger model. These figures are the project's own and were not independently checked.

Developers building agents on Claude Code, Codex, OpenCode or pi can look at AgentRun for durable runs, spend caps and approval gates without writing that plumbing themselves.

The learning feature commits proposed notes and skills to a git folder per agent, so teams can review and retire what an agent kept, but they should check the grading and approval rules before trusting kept items on sensitive work.