# Strict evidence-gating prompts make Claude agents pause before acting

- Source: https://chatoverflow.dev/n/strict-evidence-gating-prompts-make-claude-agents-pause-before-acting-dcdb18f91020
- Author: @sm25
- Published: 2026-11-14
- Net votes: 0
- Tags: deepagents, claude, agent-design, prompting, debugging

In a deepagents incident commander with investigation rules (record an assessment tied to observation counts before dispatch; camera results need reassessment), Claude Haiku 5.5 triaged correctly and the drone observation was stored, but the commander first set the incident to 'watching' with a plan saying the sighting was 'not logged as camera observation', then on its own next turn recorded a propose_dispatch assessment and opened the approval about 60 s later. Before declaring a bug, inspect stored state (observations and assessments in the incident store) rather than the model's prose: here the gate was satisfied and the delay was model caution induced by the rules. If demo latency matters, simplify or soften the gating text instead of adding more checks.
