Parcha Labs' AgentRun Teaches Coding Agents to Learn and Cut Repeat Costs Parcha Labs open-sourced AgentRun, an Apache 2.0 layer that wraps Claude Code, Codex, Pi and OpenCode to add learning, crash recovery and centralized monitoring, and the repository has reached 406 GitHub stars. AgentRun records what an agent learned from completed runs and turns repeated tasks into reusable workflows, while its new Computers Protocol gives agents a common interface to sandboxes, browsers, desktops and other remote machines. On a code-review template, AgentRun caught 42% of real bugs versus Greptile's 26% at $0.18 per review, though the results were noisier; self-hosting is free, a Render deployment costs about $27.50/month, and the project remains in alpha/beta. Parcha Labs' AgentRun Teaches Coding Agents to Learn and Cut Repeat Costs Parcha Labs open-sourced AgentRun, a layer that sits around Claude Code, Codex, Pi or OpenCode and adds learning, crash recovery and swappable computers, so repeated agent work gets cheaper. - Parcha Labs open-sourced AgentRun https://github.com/Parcha-ai/agentrun?utm source=alphasignal under Apache 2.0. It wraps Claude Code, Codex, Pi or OpenCode. - Adds learning, crash recovery and a fleet dashboard. Agents can turn repeated tasks into reusable workflows. - The new Computers Protocol asks sandbox vendors to send PRs adding their own providers. - The agent loop runs in a cheap runner and borrows a sandbox, GPU, phone or browser only per tool call. - Code review template https://agentrun.ai/blog/saving-80-percent-on-code-reviews-using-agentrun?utm source=alphasignal : 42% vs Greptile's 26% of real bugs caught, at $0.18 per review, but noisier. - Free to self-host. Render deploy about $27.50/month, or use hosted AgentRun Cloud https://console.agentrun.ai/?utm source=alphasignal . Still alpha/beta. Parcha Labs, maker of the business-research agent Grep, has released AgentRun https://github.com/Parcha-ai/agentrun?utm source=alphasignal , an Apache 2.0 platform that wraps existing coding-agent harnesses with self-improvement, durable execution and centralized monitoring. Its latest addition, the Computers Protocol, gives agents a common interface for using sandboxes, browsers, desktops and other remote machines. The repository has 406 GitHub stars. Why agents keep paying to relearn A coding agent can solve an unfamiliar task by reading documentation, inspecting files and planning a sequence of tool calls. When it receives the same class of task again, it often repeats that work. The repeated research consumes tokens, increases latency and raises the cost of every run. AgentRun targets recurring knowledge work such as document checks, business due diligence, vendor onboarding and code review. Companies may run these jobs thousands or millions of times with different inputs. Standard agent harnesses optimize the loop that reaches an answer; AgentRun also records what the agent learned and attempts to reuse it. Three layers around an existing harness AgentRun supports pi as its default harness, along with Claude Code, Codex, OpenCode and any harness that implements the Agent Client Protocol. A harness manages the loop between a model and its tools. AgentRun adds three capabilities around that loop: - Learning: The agent reviews completed runs and proposes notes, reusable skills and workflow drafts. A repeated task can eventually become a fixed workflow that skips some planning on later runs. These changes update stored knowledge and procedures while leaving the model’s weights unchanged. - Durable execution: This story is for Pro members You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.