Trace everything your agent does

Your agent said the deploy finished. The deploy did not finish. Somewhere between the plan and the receipt, a tool call failed quietly and the model paved over it with confident prose.

This Drop is about never being in that room again.

#Why polite lies happen

A language model completes text. When a tool result is ambiguous, the cheapest completion is agreement. Nothing in the loop stops the agent — unless you built the stop.

The model is not malicious. It is agreeable. Those failure modes look identical from the outside.

“Failures become diffs, not vibes.”

The fix is not a smarter prompt. The fix is evidence:

• one wide event per meaningful behavior
• a request id that threads every hop
• receipts for every write — no receipt, no claim

agent-trace-kit/├── envelope.ts          // the schema, annotated├── wrap-client.ts       // model-call choke point├── wrap-tools.ts├── replay/│   ├── harness.ts       // pin + re-run any step│   └── diff-report.ts└── fixtures/            // six real-shaped traces

#Wire the trace

#The event shape

Keep one canonical event per request-shaped unit. Wide, high-cardinality, boring names.

export const traceToolCall = (call: ToolCall, result: ToolResult) => {
  logEvent("agent.tool.completed", {
    claim: call.statedOutcome, // what the model said
    evidence: result.exitCode, // what actually happened
    request_id: call.requestId,
    tool: call.name,
  });
};

#Read it back

mega observability query --apl "['mega-dev'] | where claim != evidence"

#The audit loop

1. Pin the run

   Grab the request id from the summary the agent gave you.
2. Query the trace

   Every hop shares that id — the story assembles itself.
3. Diff claim vs evidence

   Where they disagree, you found the polite lie.

#What to log

The pattern generalizes past agents: any system that narrates its own work needs a second channel that records it.¹