Agent Work Loop Rationale

Use this reference only when reviewing or changing Agent Work Loop definitions. It records the primary-source rationale behind the five reader questions; it is not an evidence source for scoring a project and never substitutes for opened local instructions, code, tests, delivery state, or reviewed Task Episodes.

Task Understanding

OpenAI's Harness Engineering motivates specified intent, repository-local legibility, and enforceable architecture as prerequisites for reliable agent work. Agent Work Loop turns those concerns into an acceptance boundary, authoritative-context review, and explicit scope/effect boundary.

Controlled Execution

The same Harness Engineering account motivates isolated startup, agent-accessible tools, local observability, and mechanically enforced boundaries. Agent Work Loop therefore distinguishes reproducible startup, supported operation, and permission enforcement rather than treating tool inventory as behavior.

Change Validation

Google's Testing Overview motivates behavior-focused tests, controlled inputs, observable results, and explicit failure testing. The OpenTelemetry Logs specification motivates correlation across execution context, components, logs, metrics, and traces. Together they support relevant verification, attributable diagnosis and repair, and same-scope revalidation.

Reliable Delivery

GitHub's protected branch contract motivates current-revision required checks, review gates, and controlled bypass. Agent Work Loop keeps that delivery decision separate from local validation and adds risk-proportionate approval and recovery evidence.

Learning Capture

Google SRE's Postmortem Culture motivates reviewed recurrence evidence, contributing causes, owned preventive actions, discoverable knowledge, and follow-up effectiveness. Agent Work Loop therefore requires bounded opportunity detection, smallest-owner Loop Engineering, and longitudinal outcome or maintenance validation.

Interpretation boundary

These sources explain the shape of the model; they do not freeze terminology, runtime versions, provider features, or numeric scores. A reviewer must still apply the canonical Agent Work Loop checks to current project evidence and use the repository's owner chain when it is stricter or more specific.