Show HN: AgentRun – open-source, Self-improving Managed Agents

Practical AI: Tools, Models & Frameworks

What changed

AgentRun is an open-source platform that adds self-improvement, durable execution, and a control plane for agents running on multiple harnesses (Claude Code, Codex, OpenCode, pi). Agents automatically extract reusable knowledge (notes, skills, workflows) after each run, persist incomplete executions through server restarts, and expose debugging/approval controls via a unified console. It runs locally with one command, supports sandboxed execution via Docker, and costs nothing to self-host.

Why it matters

For builders running agents at scale, AgentRun directly reduces token spend by avoiding redundant reasoning—case studies show code review costs dropping from $180→$6,451→18¢/review and merchant classification from $0.89→$0.12 per item after the agent learns. The durable execution model eliminates wasted compute from crashes, and the learning system lets cheap models handle repeated patterns while expensive ones solve novel problems, making agent workflows economically viable for real-world operations.

How it compares

Unlike OpenAI's Agents API (managed, Codex-only) or AWS/GCP managed agent offerings (vendor-locked, opaque execution), AgentRun is open-source, multi-harness compatible, and gives builders full control over execution, learning, and cost via explicit approval gates and spend caps. The learning layer is novel: it systematically captures what agents discover and reuses it, rather than treating each run as isolated—turning agents into systems that improve operationally rather than requiring manual prompt tuning.

Sources