Image: GitHub
Why it mattersTeams that run agents in production can drop most of the glue code around memory, cost control and durability. The reusable pattern is agents that get cheaper as they learn their task.
A coding agent that writes a code review today starts from nothing tomorrow. It forgets the project rules, the file layout, the way the team likes a diff. If it works in an all-night batch, nobody catches it when the run stalls at 3 am. The glue code to fix all this is the work most teams end up writing themselves.
Parcha Labs and Grep.ai released AgentRun on 12 November on Show HN. The post has 51 points at the time of writing. AgentRun is an open-source platform for agents that keep notes, skills and workflows between runs. One control plane holds the spend caps, approvals and monitoring. The repository is at github.com/Parcha-ai/agentrun. The licence is Apache-2.0. The repo shows 531 stars and 1,749 commits on the main branch.
What the platform does
The team lists four pieces. The first is self-learning. An agent retains notes, skills and workflows between runs. The next run picks up where the last one left off. The second is durable execution. A server restart does not lose the agent's state. The third is a control plane. One place to set spend caps, approvals and monitoring across agents and projects. The fourth is multi-harness support. AgentRun works with Claude Code, Codex, OpenCode and the team's own pi harness.
An agent can use different environments as its computer. The release lists a sandbox, a desktop, a phone and a browser. A typed workflow option runs the deterministic parts of a task without a model. The team says this is cheaper and more reliable than letting the agent decide every step.
The numbers the team gives
The release shows two case studies on the project page, both measured by Parcha. A code-review agent moved from $6,451 a month to 18 cents a review. The team says the newer agent caught 1.6 times the bugs. A merchant classification agent moved from $0.89 a merchant to $0.12. Both numbers are the team's own, not an independent test. A reader who adopts the platform should measure on their own task.
The quickest way to try the platform is npx @parcha/agentrun, which starts a local control plane. The repository shows a one-click deploy to Render for a hosted setup.
Where it fits
AgentRun sits next to agent frameworks like LangChain, AutoGen and CrewAI. The difference the team names is the control plane and the durable state. A framework helps an agent run once. AgentRun is built for an agent that runs the same task many times. The agent gets cheaper as it learns the task. The company behind it, Grep.ai, is in the Fall 2026 Y Combinator batch.
For a team already running agents, the useful question is one. Does the control plane replace the glue code the team wrote itself? If the team runs one agent, the glue is small and the platform is heavier than the problem. If the team runs many agents across many projects, the control plane is the piece that matters. The spend caps, approvals and shared memory earn their keep then.
Source
- Parcha Labs, Parcha-ai/agentrun, the open-source repository.
- miguelrios on Hacker News, Show HN: AgentRun, open-source, Self-improving Managed Agents, the launch discussion.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.