For tasks that are too complex for a single edit and need durable progress without losing quality.
Methodology / Workflow
Agent Workflow Loop
A repeatable loop for agent work: plan, build, verify, adversarial review, anti-slop cleanup, and ship with evidence.
Plan→Build→Verify→Devil→Recheck→Ship
How it works
Turn intent into scope, edit inside that scope, verify with real commands, attack the result, remove slop, then ship only when evidence is attached.
A completed artifact plus a trace of what changed, what passed, what was reviewed, and what remains intentionally out of scope.
Operating guide
Best for tasks broad enough that a single pass will likely produce a fake green: coding work, site audits, generated artifacts, or anything that needs repeated verification.
Define the target, run one scoped build attempt, verify with project gates, audit/recheck/devil the result, repair only in scope, then ship or block with evidence.
A site-audit loop indexes pages, fixes one cluster, browser-checks it, updates the ledger, commits that cluster, and only then moves to the next cluster.
The loop fails when it becomes unbounded perfectionism, parallel writers touching the same files, or a vague “keep improving” task with no stop condition.
What you get
The minimum output is a scoped artifact, verification log, review verdict, changed paths or commit, and a clear stop/continue decision. If a blocker appears, the loop should preserve evidence rather than hide it.
Taxonomy boundary
Related shelf: `/agent-workflow/` is the system timeline; this page is only the loop rule for running scoped agent work without trusting the first green.