What is tool-call hallucination in AI agents?

When an agent claims to have called a tool and reports results without actually executing it, or calls a tool with invented parameter names and malformed arguments that crash the system. Unlike factual hallucinations, structural ones fail immediately and silently.

How common is tool-call hallucination?

AgentHallu 2026 found even top-tier models only 41.1% accurate at locating where tool hallucinations occurred. ICLR 2026 found RL-trained reasoning models increase tool hallucination rate in lockstep with task performance gains.

How does having more tools increase hallucination risk?

Research shows tool-calling hallucinations increase with tool count. Exposing all MCP tools globally pollutes the context window and confuses the LLM by presenting irrelevant tools. Focused toolsets per agent task reduce the failure rate.

← Back to dashboard

clawsmith.com/signal/ai-agent-tool-call-hallucination-wrong-params

⚠ IssueCompetitiveai_agent_mcpLive

AI agents hallucinate tool calls with wrong parameters or fake execution without calling anything

A structural hallucination pattern: agents call tools with invented parameter names, malformed arguments, or describe tool outputs they never received. The Reasoning Trap (ICLR 2026) found that RL-trained reasoning models increase tool hallucination in lockstep with task gains. AgentHallu benchmark found even top-tier models only 41.1% accurate at locating where tool hallucinations occurred.