68published protocols

AI Loop Library

Operating rule / 001

The loop must know when to stop.

ENG / 003CI repair until green KNW / 011Project docs freshness GRO / 007Monthly rank experiment

Open catalogBounded agent workflowsVol. 01 / 2026

68public loops

10work categories

04required loop parts

Freecatalog + MCP server

AI loop

noun

A bounded operating cycle where an AI system acts, checks evidence, and decides whether to continue, stop, or ask for approval.

1 Trigger2 Act3 Verify4 Exit

The evidence and exit are named before the work repeats.

Full definition

02 / Browse by work

Start with the job,
not the tool.

Every pattern includes a trigger, allowed actions, verifier, stop condition, and approval boundary. Pick a lane, then adapt the protocol to your system.

AEngineering21 loops

Ship code that survives contact with reality.

CI repair, browser verification, repo readiness, flaky tests, and scoped release loops.

Explore engineering BContent + Growth13 loops

Move from source to proof to publish.

Claim checks, buyer research, rank experiments, and learning loops.

Explore content COperations10 loops

Keep recurring work moving.

Handoffs, production sweeps, tool health, and milestone control.

Explore operations DEvaluation + Safety10 loops

Replace "looks good" with evidence.

Adversarial review, regression checks, approval gates, and red teams.

Explore evaluation EKnowledge4 loops

Capture knowledge that stays useful.

Research artifacts, ingestion QA, memory cleanup, and freshness.

Explore knowledge FPersonal Ops5 loops

Close invisible obligations.

Inbox triage, calendar promises, follow-ups, refunds, and renewals.

Explore personal ops

Open the complete library

03 / See the difference

One round.
Visible proof.

Run a small local simulation to see how an agent moves from a trigger to a defensible stop.

A browser-only demonstration. It does not call an agent or touch a repository.

stateidle

round0

verifierpending

stoppending

Awaiting trigger...

04 / Agent access

Give your agent the field manual.

The read-only MCP server can browse the catalog, shortlist patterns for a goal, critique a draft loop, and render a runnable protocol.

Install the MCP Open the skill file

Third-party quality check

Install / local stdio

mkdir -p ~/.ai-loop-library
curl -fsSL https://ailooplibrary.com/mcp/server.py \
  -o ~/.ai-loop-library/server.py
claude mcp add ai-loop-library -- \
  python3 ~/.ai-loop-library/server.py

Read-only · stdlib only · no account required

05 / Built to be referenced

Readable by people.
Legible to machines.

The library exposes plain definitions, stable loop pages, structured data, and raw source files. No signup wall stands between a reader and the evidence.

01ResearchAnalysis of 42 real-world agent loops.02GlossaryShort definitions for loop engineering terms.03Catalog JSONThe complete machine-readable loop catalog.04llms.txtA compact map for AI retrieval systems.

FIELD NOTE / 006

A good agent needs more than a prompt.

Give it a boundary, a verifier, and an exit.

Find your first loop