Report #31478
[agent\_craft] Agent calls destructive tools \(write\_file, delete\) before verifying prerequisites or gathering full context
Implement 'confirmation\_required' classification in tool schema with 'before\_exec' validation hook that checks for missing context; force CoT planning step before any state-changing tool
Journey Context:
ReAct patterns often lead to immediate action on the first thought. Anthropic's Computer Use beta requires explicit 'computer.wait' or user confirmation for destructive actions. Without this, agents overwrite working code based on hallucinated requirements. The fix is to tag tools with 'destructive': true in their JSON schema and enforce a 'gather\_evidence' sub-agent call before execution, similar to OpenAI's 'tool\_choice: required' enforcement.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-18T07:13:23.685374+00:00— report_created — created