Submit

AgentRun ships as open-source self-improving agent platform

Grep.ai open-sourced AgentRun, an Apache-2.0 platform adding self-improving skills and durable runs to Claude Code, Codex, Pi and OpenCode agents.

Written by AiAgentsListing Team

AgentRun ships as open-source self-improving agent platform

What happened

Grep.ai, the team behind the Grep.ai business-agent platform, open-sourced AgentRun on 12 November 2026, a management layer that adds self-improvement, durable execution and a control plane on top of an existing agent harness. The release is licensed Apache-2.0 and works with Pi, Claude Code, Codex and OpenCode. The "Show HN" post reached 89 points and 23 comments on Hacker News within hours of going live.

AgentRun is an open-source platform from Grep.ai that sits on top of an agent harness like Claude Code, Codex, Pi or OpenCode. It adds three things a harness does not provide on its own: agents that turn repeated steps into reusable skills and workflows, runs that survive a server restart, and a console with approvals and spend caps.

What changed

AgentRun adds self-improvement, durable execution and a control plane around whatever harness an agent already runs on.

  • Self-improvement: when an agent's definition sets "learning": true, it gets one extra turn at the end of each run to propose what to keep, a short markdown note, a reusable skill, or a workflow draft. Code checks the proposal first, then one grader judges it, and accepted items commit to the agent's own git folder on a single "main" branch.
  • Once a learned skill has served three runs, AgentRun asks the next run to turn it into a workflow draft, a JSON document of typed steps that executes as one tool call rather than a full agent loop. Drafts are dry-run before they are kept.
  • Durable execution: every agent, revision, run and step lives in one store, SQLite by default or Postgres with AGENTRUN_STORE_URL. Stopping the server mid-run and restarting it lets the run carry on; a dead runner's claimed runs are picked up by another runner once a 30-second lease expires.
  • Control plane: each tool a run can call is set to allowed, ask-first or denied; a run can be given a spend cap that pauses it instead of letting it fail, and an agent can stop mid-run to ask a person a question.
  • Agents can run with no compute, a Docker, Daytona, Modal or E2B sandbox, an Orgo or E2B desktop, a rented Android phone, or a Browserbase or Kernel browser.
  • According to the Hacker News submission text, AgentRun also works with any other harness that supports ACP, the Agent Client Protocol.
  • The release ships five ready-made templates (code-review, kyb, market-monitor, inbound-lead, release-notes) and documents two in-house results: a code-review agent that replaced a $6,451-a-month Greptile bill with an 18-cents-a-review in-house reviewer that caught 1.6x the bugs on 30 unseen PRs, and a merchant-classification job whose cost per merchant fell from $0.89 to $0.12.

Why it matters for agent builders

AgentRun matters to agent builders because it covers infrastructure most of them end up building by hand anyway: a store that survives a crash, a per-run spend cap, and a way to turn a repeated job into something cheaper than re-running the same reasoning loop. Most agent harnesses, including Claude Code, Codex and OpenCode, handle reasoning and tool calls but leave run storage, recovery and cost control to whoever deploys them. AgentRun's own code-review case study targets teams past the prototype stage: a review bill rose 36x in a single year once agents started reviewing pull requests by default, before the team built a cheaper in-house reviewer on top of AgentRun's learning.

What to watch

The open question is whether AgentRun's spend caps actually bound everything an agent can spend, including work it offloads elsewhere. A commenter on the Hacker News thread asked whether caps also cover detached background processes an agent spins up, and a Grep.ai team member posting as miguelrios replied only "I will take a look into this," leaving the question open at launch. The homepage's own comparison table, checked 7 November 2026, lists OpenAI's Agents API as "not documented" for learning, which means AgentRun's comparison against Anthropic's Claude Managed Agents, OpenAI's Agents API and Google Cloud's Gemini Enterprise Agent Platform is its own assessment, not a verified benchmark.

Key takeaways

  • AgentRun is Apache-2.0 and was open-sourced by Grep.ai (YC F26) on 12 November 2026; it layers on top of Pi, Claude Code, Codex and OpenCode rather than replacing them.
  • Learning turns repeated steps into markdown notes, reusable skills and workflow drafts, checked by code and a grader before being committed to the agent's own git folder.
  • One in-house case study replaced a $6,451-a-month Greptile bill with an 18-cents-a-review in-house reviewer; another cut merchant classification from $0.89 to $0.12 a merchant.
  • Every run, revision and step lives in one store (SQLite or Postgres), so a crashed server can restart and resume a run instead of starting over.
  • Spend caps pause a run rather than fail it, but whether they cover an agent's detached background processes went unanswered at launch.

FAQ

What is AgentRun?

AgentRun is an open-source, Apache-2.0-licensed platform from Grep.ai that adds self-improvement, durable execution and a control plane on top of an agent harness such as Claude Code, Codex, Pi or OpenCode, rather than acting as a harness itself.

Which agent harnesses does AgentRun support?

AgentRun works with Pi (the default), Claude Code, Codex and OpenCode, named per agent with a "harness" field. Codex and Claude Code install themselves on first use, and OpenCode needs its opencode-ai package installed alongside the server.

How does AgentRun's self-improvement feature cut costs?

An agent with learning enabled gets an extra turn after each run to propose a note, a skill or a workflow draft. Once checked and accepted, later runs reuse that instead of re-planning the same steps. Grep.ai's own examples show a code-review agent that fell to 18 cents a review and a merchant-classification job that fell from $0.89 to $0.12 per merchant.

Source: Parcha-ai/agentrun on GitHub Discussion: Hacker News

Compare AgentRun against other coding agents built on Claude Code in the AI Agents Listing agents directory.

Share:

Subscribe to our newsletter

One email a week. New agents, MCP servers and skills, and what is actually getting traction.

More news

  1. ClawDOS ships as an open source AI coding agent for DOS

    ClawDOS is an open source AI coding agent harness running Claude's Messages API on a real 1980s IBM PC, posted to Hacker News on 21 November 2026.

  2. Turbo-Browser-MCP ships as an 11-tool Playwright MCP server

    Turbo-browser-mcp is a new open-source Playwright MCP server with 11 tools that batches browser actions to cut token use and model turns for AI agents.

  3. Windows 11 diagnostics flag Claude Code, Codex and Cursor

    A Windows 11 component called MissionControlAggregator checks for Claude Code, Codex, Cursor and other coding agents, researcher Xusheng Li found on 16 November 2026.