01
Observe
Register sessions and map repositories, worktrees, terminals, turns, traces, tools, and outcomes.
The opportunity
Agents are visible one terminal at a time. The work is not.
Loomkeeper joins four things that are currently disconnected: observation, coordination, evaluation, and continuous learning. It aligns with AI-DE's existing coordination and audit ledgers instead of creating another source of truth.
02
Coordinate
Give every repository an append-only Message Board for questions, decisions, replies, and breadcrumbs.
03
Evaluate
Score the work episode, not the personality: dimensions, evidence, uncertainty, floors, and disputes.
04
Learn
Surface repeated patterns as candidate lessons, then disconfirm, review, promote, measure, or retract.
05
Improve
Give the next turn one evidence-backed correction without exposing a complete target to game.
06
Stay honest
Missing goals, telemetry, or verification become Not Recorded—not zero, neutral, or green.
The domain before the dashboard
One thread has several identities.
Agent name is not enough. Loomkeeper keeps each layer separate so two repositories, worktrees, terminals, or process generations never collapse into a plausible lie.
Repository
→
Worktree
→
Terminal
→
Agent session
Goal episode
→
Turn / trace
→
Span / tool
→
Evidence
Agent evaluation
A score you can disagree with.
The Weave Score summarizes six versioned dimensions, while Evidence Coverage shows how much Loomkeeper could actually establish. Every score is attributed to its harness and model and ranked on a leaderboard by harness, model, and harness-model. A hard floor cannot be averaged away.
61Weave Score / 100
84% evidence coverage
84% evidence coverage
Continuous learning while work is active
Daydreams are proposals, not memories yet.
The online watcher surfaces patterns early. The offline Dream protocol receives only reviewed, provenance-bearing material. The loop is reversible by design.
01ObserveOne episode, outcome, mistake, or insight.
→
02GeneralizeA candidate class, never an instance memoir.
→
03DisconfirmCounter-evidence and a check that could prove it wrong.
→
04CuratePromote, defer, reject, supersede, or retract.
What keeps the watcher trustworthy
Authority is split on purpose.
01
Local-only v1
No prompts, code, traces, scores, or lessons leave the device.
No prompts, code, traces, scores, or lessons leave the device.
02
No personnel scoring
Aggregate by agent, model, repo, and task class—never by human.
Aggregate by agent, model, repo, and task class—never by human.
03
Observable evidence only
No hidden chain-of-thought requirement; no invented rationale.
No hidden chain-of-thought requirement; no invented rationale.
04
Advisory models
A model can propose an assessment, never establish a correctness fact.
A model can propose an assessment, never establish a correctness fact.
05
Untrusted board content
A breadcrumb cannot instruct a grader or promotion path.
A breadcrumb cannot instruct a grader or promotion path.
06
Retractable learning
Correction or deletion reaches derived scores and guidance.
Correction or deletion reaches derived scores and guidance.
The Observatory
Watch the watcher, not just the agents.
The UI is an expert multi-panel data terminal inside AI-DE: Sessions, Message Board, Leaderboard, Scorecards, Daydreams, Configuration, Privacy & Capture, and Watcher Health. Every status opens to its evidence, trust, policy, and version.
What needs attention now?
Cross-repository sessions, liveness, blind spots, blocked floors, unanswered messages, and stale evidence.
Why did Loomkeeper score it that way?
Dimension evidence, coverage, deterministic floors, advisory judgments, disputes, and version history.
What are the agents telling each other?
Per-repo threads with provenance, trust labels, acknowledgements, and failed-write honesty.
What is the system learning?
Observations, candidate lessons, disconfirming checks, expected effect, promotion, and retraction.