Catch Your Agent Inventing Tool Results (Before Your Users Do)

Описание и применение

Проверка фактического выполнения инструментов агентом через исполнительный реестр позволяет выявлять вымышленные или неверно цитируемые результаты, повышая надёжность автоматизированных рабочих процессов.

Задача

AI agents can fabricate or misquote tool results, leading to hallucinated outcomes that pass transcript reviews.

Результат

Auditing against an executor-side ledger reliably uncovers fabricated or misquoted tool results, ensuring agents only report genuine executions.

Рабочий процесс

• 1. Record every tool dispatch in a ledger: tool_call_id → {tool, args, raw result}.
• 2. After agent run, extract all factual claims about tool outcomes, noting referenced call IDs and asserted results.
• 3. For each claim, check ledger: if call ID missing → FABRICATED; if result mismatches → MISQUOTED; else OK.
• 4. Flag any FABRICATED or MISQUOTED findings for review.
• 5. Repeat audit after every run that touches real resources.

Доказательства

Audit them against an executor-side ledger.

Источник

Your executor records every dispatch: tool_call_id → {tool, args, raw result} in a ledger the agent can't write to.

Источник