HAR – open-source harness for multi-agent coding workflows

I'll pull the GitHub page for concrete project details so the paragraphs stay factual and specific.**HAR** (github.com/os-factory/har) is an open-source, agent-agnostic harness for multi-agent coding workflows. It lets you run a fleet of coding agents in parallel on any repository, with deterministic validation gates, verifiable proof of what ran, and full observability across agents. The project targets Claude Code, Cursor, Codex, and any MCP-connected agent. On Hacker News it landed with 2 points and 0 comments; the package installs as @osfactory/har via npm.

Architecturally, HAR is both a CLI and an MCP server. It scaffolds a machine-readable **.har/** contract in the repo—stack, scripts, and checks in one place instead of README, CLAUDE.md, Cursor rules, and CI yaml drifting apart. A run follows Discover → Isolate → Build → Verify → Hand off. Isolate gives each agent its own **slot**: separate git worktree, branch, ports, and (when needed) database so concurrent agents do not collide on shared state. Verify runs the project’s real checks the same way every time; full verify ties a pass to a validated tree hash plus logs and artifacts. Mission Control is the local dashboard for repositories, worktrees, runs, validations, and artifacts. Plugins (e.g. Playwright, RocketSim) register extra verification stages. `har env maintain` diffs the installed harness against current templates and flags drift before silent failures.

For engineers already running multiple coding agents, the bottleneck is rarely the model—it is collision on ports and git state, re-running checks because you cannot trust an agent’s self-report, and lock-in when verification lives only inside one vendor’s sandbox. HAR moves coordination into the repo: agents edit only their isolated copy; reviewers get a branch plus evidence instead of “tests passed” claims; the contract stays portable when you switch tools. That is the difference between a demo multi-agent setup and something you can hand a reviewer.

Competitive context is crowded on the agent side (Claude Code, Cursor, Codex, and other MCP clients) and thin on open, agent-agnostic isolation and proof. Vendor harnesses optimize for one product surface; hand-rolled scripts rot when deps, services, or env vars change. HAR’s bet is an open Apache 2.0 standard in-repo, not a hosted dashboard you rebuild when you change agents. Kerno sponsors the project; docs and concepts live at harproject.dev. Early HN traction (2 points, no discussion) marks it as early-stage relative to the larger coding-agent ecosystem.

Practical next steps: try `har env init`, then `har env launch 1` and `har env verify 1 --full` on a real app with a non-trivial test suite and a local database dependency—that exercises isolation and evidence, not just happy-path lint. Watch whether the plugin ecosystem covers your stack, how commit-gate enforcement behaves in PR review, and whether Mission Control becomes the default ops view when more than one agent is running. If your team already juggles multiple agents on one monorepo, the slot model and tree-hash verification are the pieces worth evaluating first.