Reduce agent hallucination by giving sub-agents fewer tools and read-only verification scopes

Digital Applied·medium signal

Spawning sub-agents with role-restricted tool sets (Explore/Plan/Verify) enforced server-side prevents research and planning agents from accidentally writing, shrinks the prompt-injection surface, and — counterintuitively — reduces hallucination by removing tool-choice ambiguity. Anthropic's isolated multi-agent setup reported a 90.2% improvement over single-agent and 84% token reduction, largely because each sub-agent works in a fresh context and returns only a summary. Scope tools to the role, not the maximum.