Loom

Loom is the uniform, verifiable contract for running agents — any agent, any runtime. The manifest is the shared interface; the value is the verification and audit that makes non-deterministic agent work trustworthy and controllable at scale.

Loom deploys a coding agent as a repeatable run. Instead of babysitting an agent in a terminal, you write a Weave — a small JSON file that declares the sandbox to run in, the goal, what "done" means, and what to report back. Loom spins the sandbox, runs the agent, checks the result, and ships it as a branch and a pull request.

Think of a Weave as a manifest for an agent run — the way a Kubernetes manifest describes a deployment, a Weave describes an agentic task.

A Weave

{
  "spec": "nautloom/v1",
  "kind": "Weave",
  "metadata": { "name": "build-verify", "version": 1 },
  "runtime": {
    "provider": "gitvm",
    "template": "agent-desktop",
    "resources": { "vcpus": 4, "memoryMB": 8192, "ttl": 21600 }
  },
  "intent":     { "goal": "…what the agent should do…" },
  "acceptance": [ /* checks that define done */ ],
  "report":     { "include": ["summary", "tests", "video"] }
}

runtime says where it runs, intent is the task, acceptance is what makes it a pass, and report is what you get back. Real examples — blank, quick-run, build-verify, ui-record, swarm — live in the repo.

How a run works

1. Spin an isolated sandbox — a GitVM microVM, fresh per run.
2. Sync the working directory and goal into it.
3. Run the agent (Claude Code or Codex) — live-streamed, secret-redacted log.
4. Check acceptance — tests and criteria.
5. Pull the code back and ship it: a branch and a PR.
6. Report: a summary, test results, and a session recording.

The runtime — GitVM

Weaves run on GitVM, a Firecracker microVM sandbox substrate. The default agent-desktop template is an Ubuntu desktop with a screen recorder and the agent CLIs — so every run is isolated, watchable, and recorded. One sandbox per working directory; parallel runs each get their own git worktree.

Status