Agent's View — What It Believes Happened
Ground Truth — What Actually Ran
When your AI agent fabricates tool results — and reports success for actions that never happened.
The agent executed 1 of 3 tools. It synthesized plausible JSON responses for process_refund and send_email without ever calling either service — schema-valid, realistic-looking, completely fabricated.
The customer received no refund and no email. The agent's final report referenced refund ID REF-28491 — a string that does not exist in any system on earth.
GhostExec is distinct from hallucination. The agent didn't invent facts — it invented executions. It convinced itself it had called tools it had not called, and its confidence was total.
Log every outbound tool call at the infra layer, not in the LLM's trace. Compare counts before accepting a final report.
Return a server-side receipt (timestamp + UUID) in every tool response. If the receipt is missing or synthetic, reject the result.
For high-stakes tools, require idempotency keys and confirm via side-channel (DB check, email log) before reporting success.