Agent Loop Reliability

This skill is derived from: https://claude.com/blog/building-agents-with-the-claude-agent-sdk (published 2025-09-29).

Instructions

• Start every run by writing down the current objective and success criteria.
• Gather context with explicit actions (file search, logs, docs) before making changes.
• Take action in small, reversible steps; prefer tool calls or scripts over manual reasoning for deterministic work.
• Verify each step using one or more of: deterministic rules, visual validation, or an LLM-as-judge rubric.
• If the agent is stuck, change the environment (better tools, better search, stricter rules) rather than repeating the same attempt.
• After completion, capture failures as test cases and add them to an eval set.

Examples

Example: turn an unreliable task into an agent loop

1. Gather context: search logs for the failing module, open relevant files.

2. Take action: apply the smallest patch.

3. Verify: run unit tests and lint.

4. Repeat until clean.

Example: verification checklist

Use references/verification-checklist.md as a standard review rubric for UI or content output.

Templates

See:

• templates/agent-loop.md
• templates/verification-plan.md

References

See:

• references/verification-checklist.md
• references/testing-improvement-questions.md