What is Loom
Every CI tool can tell you a run failed. Loom tells you exactly where, in three hops.
Loom is an open-source workflow engine that writes deterministic artifacts as it runs — receipts, manifests, and event streams — so when something breaks, you follow pointers straight to the failing unit. No scrolling. No grepping. No guessing.
Your AI agent can do the same thing: read structured JSON, follow the same pointer chain, and land on the same evidence you would. No log parsing. No token-heavy context windows.
How Loom works
loom check catches schema and logic errors before you ever run.loom run --local executes your workflow and writes structured artifacts.What makes Loom different
Most workflow tools treat output as a stream — text flows by, and when something breaks, you scroll backward through it hoping to spot the signal. Loom takes a different approach: every run produces deterministic artifacts organized for navigation, not just recording.
The core idea is a pointer chain. A receipt points to a pipeline manifest, which points to individual job manifests, which point to granular event streams. You never read everything — you follow the chain to the narrow slice that matters.
This is not just a nicer log viewer. It is a fundamentally different diagnostic model — one that works the same way whether a human is reading the artifacts or an AI agent is parsing them programmatically. The artifacts are small, structured, and self-describing. The paths are deterministic. The evidence is reproducible across runs.
Built for you AND your agent
For you
- Three hops to the failure — follow receipt → manifest → events instead of scrolling thousands of lines.
- Reproducible evidence — same run, same artifacts, same paths. Hand a receipt to a teammate and they see what you see.
- Faster iteration — re-run, diff artifacts, confirm whether the fix landed. No guesswork.
For your agent
- Structured JSON — no regex, no log parsing. Agents read manifests and events directly.
- Low-token footprint — a receipt + one events file is ~2 KB, not megabytes of raw output.
- Deterministic paths — agents can navigate the artifact tree without discovery heuristics.
Loom is in active Alpha 1 development. The pointer navigation model — receipt → manifest → events — is stable and designed to stay that way. Execution is local-only today, with host and Docker-backed jobs on Linux. Remote execution, broader platform targets, and shared runners are on the roadmap. This is the right time to evaluate the diagnostic model and tell us what's missing.