Prefer measurable goals
Tests, timings, coverage, screenshots, retrieval checks, source tables, and explicit rubrics beat “looks good.”
Agent workflows, minus the cult brochure
Loops are what come after prompts.
A practical field guide and searchable library for bounded AI operating cycles: trigger, objective, repeated action, verification, stop condition, and safe output.
What this is
A loop is a bounded operating cycle for an agent.
Prompts ask once. Agents complete one task. Loops keep operating until the goal, budget, blocker, or approval boundary is reached.
They work best where the job is repetitive, evidence-backed, and annoying enough that a human keeps putting it off: CI speed, docs freshness, stale knowledge, source QA, product checks, pre-publish fact review, dependency triage, and customer follow-up.
They fail when the goal is vague, the stop condition is emotional, or the agent can mutate production without a leash. The technical term is “bad idea.”
One bottleneck at a time
A loop should make one coherent move, verify it, keep or revert, then continue.
State survives the run
Long work needs progress files, evidence logs, branches, or notes. Chat memory is not infrastructure.
Humans own judgment
Publishing, sending, deleting, shipping, money, legal, and reputation moves stay approval-gated.
Library
Copyable loops for coding and non-coding work.
0 loops. Search title, use case, tags, verification, or prompt.
No loops match that filter. Either you found a gap, or search is being dramatic.
Builder
The six-part loop spec.
Trigger
Manual, scheduled, or event-based. “After each source note” beats “whenever.”
Objective
The thing to make true. Prefer measurable, bounded, and boring.
Repeated action
Inspect, plan, change, verify, record, repeat. Small moves. No heroics.
Verification
Commands, metrics, screenshots, claim tables, browser checks, or rubriced review.
Stop condition
Success, budget, blocked access, no progress, or approval required.
Safe output
PR, patch, report, draft, notification, or gated recommendation. Not surprise production changes.
Field notes
Where loops are strongest right now.
Coding loops
CI, test stabilization, dependency triage, repo readiness, docs sync, and adversarial PR review have clear evidence paths.
Knowledge loops
Source ingestion, stale memory cleanup, research-to-artifact, and pre-publish source checks turn scattered information into usable leverage.
Operations loops
Living stories, support follow-up, recovery proof, toolchain health checks, and customer AI deployment keep work moving when humans would forget.
Content loops
Weekly post experiments, buyer-objection research, SEO/GEO visibility, and social source-to-insight loops keep distribution grounded in evidence.