Agent Work Loop

The Agent Work Loop reviews whether a coding-agent Harness can carry one task from a clear intent to a validated, reliably delivered result, then turn supported learning opportunities into improvements that remain useful over time. It is not a mandatory process, transcript score, asset inventory, or repository maturity model.

The loop being scored

A good harness lets an agent finish a bounded, recoverable loop:

understand context -> make a bounded change -> choose the right validation ->
interpret failure -> repair -> re-run validation -> state residual risk

The five dimensions

Dimension	The question it answers	Backed by
Task Understanding	Does the agent know the goal and what "done" means?	Rules, AGENTS.md, specs, DESIGN.md
Controlled Execution	Is the work on supported, repeatable paths?	Skills, commands, MCP tools, sandbox boundaries
Change Validation	Is there evidence the change actually works?	Tests, lint, Hooks, observable diagnostics
Reliable Delivery	Does AI speed bypass quality checks or acceptance?	Human review, approvals, CI/CD, recovery paths
Learning Capture	Does the next task benefit from this one?	Loop Discovery, reusable SDLC Skills, Memory

Each dimension resolves three stable checks (fifteen check ids in total) before it is scored. Learning Capture is scored independently of the first four dimensions.

Review unit: the Task Episode

A Task Episode is one user goal with one acceptance boundary. It may span multiple turns or sessions, but every claim must stay tied to the same goal, target, action, and result. Unrelated work is never merged because it happened in one session, and aggregate counts never become task behavior.

When eligible session evidence is partial or unavailable, the review stays a session-limited Agent Work Loop review: unavailable behavior remains Unobserved, and inspected project Harness evidence supports only the mechanisms it can support.

Evidence states

• Present: an owned mechanism or review contract exists;
• Wired: the relevant task, trigger, or owner route can reach it;
• Exercised: a linked episode or inspection used it and retained a result;
• Outcome-supported: a comparable later result supports the claimed effect;
• Missing: inspected evidence confirms a required mechanism or result is absent;
• Unobserved: the available observation boundary cannot decide;
• Not applicable: inspected task and project evidence proves it does not apply.

Evidence state is not pass/fail. An exercised operation may expose a defect, a safe denial may be correct behavior, and an unavailable external boundary is Unobserved, not automatically Missing.

Evidence limits score confidence

For the four current-task dimensions, the highest supported evidence state sets an absolute score ceiling:

Highest supported evidence	Absolute score ceiling
Missing, Unobserved, or Not applicable	59
Present	74
Wired	84
Exercised	94
Outcome-supported	100

These are ceilings, not a score formula. A configured capability is not observed use, and only an Outcome-supported later comparison permits a claim that an intervention improved later work.

Full model

The complete contract — check tables, findings, scoring boundaries, and longitudinal validation — lives in models/agent-work-loop.md.