AI Coding Agent Harness for Reviewed Engineering Work
Run one or several AI coding agents inside a Mrrlin harness: repository context, isolated git worktrees, checkpoints, reviews, and deploy control.
Chat tools vs Mrrlin
The execution layer beats another blank chat window.
Chat tools make you re-send context and manage the work. Mrrlin keeps project memory and spends tokens deliberately.
Why it matters
Coding agents need walls, a shared spec and proof to hand over.
Mrrlin helps coding agents work inside a reviewed harness: tasks carry repository context, worktrees isolate changes, checks create evidence, and deploy-sensitive steps wait for approval.
Workflow
From a repository task to a pull request a reviewer can trust.
Attach the repository context
The task carries the spec, acceptance checks and the repository facts that matter for it. Claude Code reads the repo and docs before the plan is final, and open questions about expected behavior go to the inbox instead of being settled by assumption halfway through the change.
One worktree per agent
Every run gets an isolated git worktree and its own branch, with the base commit recorded. Three agents working toward one goal therefore edit three separate trees, and your checkout stays exactly as you left it while they work.
Turn checks into artifacts
Build output, test results, command output and a diff summary are recorded on the run as they happen. The run's self-review is kept as well, but the verdict belongs to a second model from another vendor, which reads that evidence rather than the agent's account of it.
Hand off at the pull request
When review passes, the run moves to pull request status on your repository. The task's deploy policy decides whether it stops there, goes to a preview, or heads toward production, and production deploys run through your own GitHub Actions workflows only after you approve.
Use cases
Where it fits.
Repository implementation
Ask for the landing page hero and signup form to be rebuilt against the new spec. The AI coding agent harness hands the agent that spec, the component conventions recorded in the wiki and the acceptance checks, then records its checks and diff summary before anyone opens the branch.
Parallel agents in separate worktrees
One agent rewrites the hero copy, Claude Code rebuilds the form component, and a third verifies routes and redirects. Each works in its own worktree against the same spec in project memory, so running them in parallel does not put three agents in one directory.
Build and test checkpoints
Midway through the form rebuild, a checkpoint records a failing validation test. The reviewer returns the run with that output attached, and the resumed run carries only the delta since the last attempt instead of replaying the whole conversation from the start.
Cross-agent handoff to a pull request
The route-check task is blocked by the implementation task, so it starts only once the form change exists. It reads that run's diff summary and handoff notes, adds its own verification, and the reviewer sees every agent's evidence on one trail before the pull request opens.
Comparison
What changes when coding agents stop sharing your working tree.
Keep reading
More on coordinating coding agents.
Start from the AI coding agent orchestration overview, or go deeper with the guides next to this one.
FAQ
Before you start.
Can one AI coding agent harness run several agents on the same goal at once?
Yes. Mrrlin splits the goal into tasks, gives each running agent its own worktree and branch, and has all of them read one spec from project memory. Blocked-by dependencies order the handoffs, so a verification agent waits for the implementation it checks, and the Autopilot policy caps how many runs execute concurrently.
Does the harness replace Claude Code or Codex?
No, it runs them. Claude Code and Codex CLI start on your own computer under your existing subscriptions, and they can read and update their tasks through the Mrrlin MCP server. What the AI coding agent harness adds is what the agents do not keep on their own: scope, isolation, evidence, review and approval.
What happens when a coding agent's change fails review?
It goes back to the agent with the reviewer's notes instead of landing in your inbox as finished. The task stays open, the failed attempt keeps its artifacts, and a resumed run sends only what changed since the last try. If Autopilot sees the same task fail repeatedly, it stops instead of retrying indefinitely.
Tell us the outcome you want AI to execute.
Share the workflow you want to automate. We’ll map the first Mrrlin run — plan, agent routing, review loops, and approval checkpoints.
No credit card · No migration · One goal