Loop engineering foundry

Design the system that keeps prompting the agent

Loop engineering moves the design problem above a single prompt. It specifies how an agent observes state, decides, acts, inspects the result, records evidence, recovers, verifies, and stops across repeated model calls.

Compare ReAct, plan-and-replan, evaluator-optimizer, capped retry, human approval, and checkpointed-worker patterns across research, coding, support, and structured-extraction scenarios. Inject a failure, lower a budget, or deny approval and inspect the exact terminal state instead of accepting a false claim of completion.

An observable agent state machine with branches for tools, retries, checkpoints, approval, verification, and stopping
Loop engineering · Design iteration, recovery, and explicit terminal states
Manipulate
Pattern · scenario · failure · approval · budgets
Observe
State graph · ledger · calls · retries · checkpoints
Terminal states
Proven done · partial · awaiting approval · exhausted · failed closed
Loop pattern switchboardReAct tool-use loop
Conference research

Verify six upcoming conference dates and create a sourced briefing.

Done conditionEvery date and location resolves to an official organizer page.

01observe
02decide
03act
04inspect
05verify
06stop
Terminal stateobserve in progress

Required artifact: Conference evidence table.

Model calls1
Tool calls0
Retries0
Checkpoints0
Execution ledgerobservable state, not hidden confidence
01 · observeRead goal and evidence ledger

Required artifact: Conference evidence table.

M 1 · T 0 · R 0
Try the same loop on another job

Primary research

Inspect the mechanism at its source.

Project architecture and terminology were checked against official documentation on August 10, 2026. Research papers are linked to their original publication records.

Current definition and practiceIBM: What is loop engineering?Original source ↗Interleaved reasoning and actionReActOriginal source ↗Intake through persistenceOpenClaw agent loopOriginal source ↗Budgets, tools, compression, fallbackHermes Agent loop internalsOriginal source ↗

Keep experimenting

Change the layer, keep the system visible.

AI Lab indexSee every interactive laboratoryBrowse the suite →Transformer microscopeGo beneath the agent runtime into the modelInspect the matrices →