Agent Debugger
Skill for diagnosing and resolving failures in AI agents, tool calls and pipelines.
When to use this skill
- The agent enters a loop and never finishes
- A tool call fails silently or returns an unexpected result
- The agent does something different from what was requested
- There is a parsing error in the model response
- The context fills up and the agent loses information
- Two subagents conflict or overwrite each other
Patterns and solutions
Infinite loop — add an explicit stop condition, limit iterations
Malformed tool call — validate schema before executing, add examples in the prompt
Context overflow — summarize history every N steps, truncate tool outputs to the minimum needed
Parsing error — use structured outputs if supported, add validation with retry logic
Prompt drift — re-inject the original objective every N steps as a system reminder
Silent subagent conflict — assign exclusive resources to each subagent, implement locks
Verification checklist
Compatibility
Claude Code, Cursor, Windsurf, Cline, Roo and any agent compatible with the Agent Skills standard.