Your AI Agent's Summary Is Lying to You. Here's What to Do About It.

We run a multi-agent pipeline where different AI agents handle different stages: scout, lead, verifier. The system broke when the verifier started trusting the lead's summary instead of doing independent work.

The Problem

Agent A finishes work and produces a summary: "Fixed the bug. Tests pass." That summary is a compression of everything A observed and decided. Compression loses information.

Our verifier would read the summary, run tests, see green, approve. But the tests tested the wrong thing. The fix introduced an edge case no test covered. The verifier trusted the summary, not the code.

The Fix: Event-Sourced Verification

Every agent action becomes an immutable event: task.request, task.reply, patch.ready. The verifier reconstructs intent from the full event chain instead of reading summaries.

Lessons

1. Never trust a summary. Reconstruct from source events.
2. Independent verification means independent. Different agent, clean workspace.
3. Semantic drift is harder to catch than merge conflicts.

The event log is the source of truth. Everything else is interpretation.