awesome-loop-engineering / ANTI-PATTERNS.md
cy0307's picture
Sync awesome-loop-engineering
9ec4919 verified
|
Raw
History Blame Contribute Delete
3.18 kB

Loop Engineering Anti-Patterns

Use this page to reject unsafe, vague, or misleading loop designs.

Prompt Loop With No Contract

Symptom: A script repeatedly asks an agent to "keep improving" without a named objective, intake, state, verification, or exit condition.

Why it fails: The loop becomes a manual prompting habit disguised as automation.

Better: Define the loop contract before running: objective, trigger, intake, workspace, context, delegation, verification, state, budget, escalation, and exit.

Infinite Retry Loop

Symptom: The loop keeps feeding failures back to the agent indefinitely.

Why it fails: Cost grows, context degrades, and the loop may repeatedly damage the same files.

Better: Set retry budgets per blocker, stop on repeated failures, and escalate with evidence.

Model Self-Approval

Symptom: The same agent that made the change decides that the work is complete.

Why it fails: Models can rationalize their own mistakes and miss regressions they introduced.

Better: Use deterministic checks, a separate verifier agent, or human review for completion gates.

Hidden State

Symptom: The only memory is inside the conversation.

Why it fails: The next run cannot reliably know what was tried, what passed, what failed, or what remains.

Better: Persist state in files, issues, comments, checkpoints, traces, or workflow state.

Unsafe Production Autonomy

Symptom: A loop can deploy, roll back, change config, rotate secrets, or modify data without explicit approval boundaries.

Why it fails: A small model error can become an operational incident.

Better: Make sensitive actions approval-gated and document rollback, escalation, and owner rules.

Generic Automation Rebranded As Loop Engineering

Symptom: A cron job or workflow runs without agentic reasoning, context loading, verification, or stateful next-action decisions.

Why it fails: It blurs Loop Engineering with ordinary automation.

Better: Call ordinary automation ordinary automation. Use Loop Engineering when repeated agent work is actually being triggered, bounded, verified, persisted, and rerun.

Context Dump Loop

Symptom: Each run gives the agent a huge context bundle instead of a curated state contract.

Why it fails: The loop becomes slow, expensive, and hard to debug.

Better: Separate durable state from supporting context. Load only the state, docs, examples, and receipts needed for the current objective.

No Escalation Path

Symptom: The loop keeps acting even when it lacks credentials, sees ambiguous product requirements, hits flaky infrastructure, or needs architectural judgment.

Why it fails: The agent spends budget where a human decision is required.

Better: Define escalation triggers and include the evidence a human needs to take over.

No Receipts

Symptom: The loop says it is done without commands, logs, changed files, trace IDs, dashboard snapshots, reviewer decisions, or links.

Why it fails: Maintainers cannot audit the loop or trust its result.

Better: Require receipts as part of the exit gate.