Show HN: AgentRun – open-source, Self-improving Managed Agents
Reported by Hacker News1 source ↓
AgentRun is an open-source platform for running agents on Claude Code, Codex, OpenCode or pi, with persistent learning, workflow tools and run management. Its maker says its own code reviewer costs 18 cents per review.
Key points
- AgentRun is released under the Apache-2.0 license
- The platform can run locally, in a container or as a server with runners
- Its maker says its reviewer caught 1.6 times as many real bugs as Greptile on 30 PRs
- Docker enables a shell sandbox; without Docker, agents can read and write files but have no shell
Details
AgentRun adds run tracking, recovery, persistent learning and a control plane around supported agent harnesses. Users can enable learning for an agent; after a run, the system checks and grades proposed notes before saving approved items in a Git folder. Users can review or retire those items in the console.
The platform also supports typed workflows that combine tool calls, model steps, branching and agent handoffs.
AgentRun's maker says its custom code reviewer costs 18 cents per review, compared with a dollar per review for Greptile, and caught 1.6 times as many real bugs across 30 PRs. The maker also says a merchant-classification task cost 12 cents per merchant after transferring what a more expensive model learned, down from 89 cents. These are the maker's reported results.
Written from coverage as of
Images
- Hacker News