Parcha open-sources the AgentRun agent platform, adding learning and persistent execution to Claude Code, Codex, and more
Parcha has open-sourced the AgentRun platform, which adds self-improvement, persistent execution, and a control plane on top of agent frameworks such as Claude Code, Codex, OpenCode, and pi, under the Apache-2.0 license. When learning is enabled, at the end of each run it proposes notes, skills, or workflows, which after code validation and scoring are committed to the agent's own git directory, and the next run starts from there. The platform also provides typed workflows, approvals and spending caps, and four types of computers—sandbox/desktop/phone/browser—and can be launched locally with a single command, npx @parcha/agentrun.
Why read this
AI score 70
The original article fully explains the design trade-offs of the learning mechanism, persistent execution, and control plane, which can be used to judge which types of existing agent workflows it is suitable for integrating into.