Before: Log archaeology
Noise-heavy… 5,000 lines of interleaved stdout …
error: command failed (exit code 1)
… which job? which step? grep and hope.Interleaved text output forces manual hunting through massive logs.
ALPHA 1
Know exactly what failed — in 3 files, not 5,000 lines.
Loom writes structured manifests and event streams so agents and teams can jump to the failing unit without scanning logs. Fewer tokens consumed. Faster feedback loops.
3 files
to locate any failure
~2 KB
total manifest size
5 min
to first workflow run
Traditional CI logs vs Loom pointer-first artifacts
After: 3 pointer hops
Signal-firstpipeline/manifest.json -> failing_job_id
jobs/<id>/manifest.json -> failing_step_events_path
.../events.jsonl -> exact failure, nothing elseRead 3 files (~2 KB), jump to the failing unit, and move straight to root cause.
Built for agents
AI agents burn tokens on noise. Loom gives them signal.
Tiny manifests, not megabytes
Agents read small JSON manifests and narrow JSONL event streams — typically ~2 KB to reach the failing unit. No 5 MB log dumps eating your context window.
Deterministic pointers
Every manifest contains stable file paths to the next artifact. Agents follow pointers instead of parsing unstructured text — no regex, no heuristics, no hallucinated line numbers.
Parse, don't prompt
Structured JSON means agents can JSON.parse() their way to root cause. No natural-language extraction from log soup. Fewer tokens in, faster diagnosis out.
Why teams choose Loom
Token-efficient agent triage
Agents locate failures by reading 3 small JSON manifests (~2 KB total), not megabytes of interleaved logs. Structured pointers mean fewer tokens consumed and faster automated diagnosis.
Deterministic local reproduction
Same workflow, same artifacts, every run. Local execution produces stable, comparable artifacts so failures are reproducible and explainable — no 'works on my machine' mysteries.
Shareable, auditable artifacts
Receipts and manifests are file-addressable pointers you can attach to issues, diff across runs, and hand to teammates — no log archaeology required.
How it works
Author
Write a workflow in `.loom/workflow.yml`.
Validate
Run `loom check` to validate schema.
Execute
Run `loom run --local`.
Diagnose
Follow manifest pointers to the failing unit.
Start here
Three short paths depending on what you want to do next.