Stop trusting the chat log. AI agent observability isn't about watching a dashboard in real-time; it's about making an agent's work reconstructable from the outside. A chat transcript is just the agent's self-narration, written with the same blind spots that produced the work. True observability requires the execution layer to keep the record: one entry per tool call, minted at run time, carrying saved results in a form that allows external verification.

The Three Signals Translated for Agents

Traditional software observability relies on traces, metrics, and logs, but these change shape when applied to autonomous agents. Traces must become step sequences that show the order of tool calls and how each result fed the next decision. Metrics shift to cost, latency, and failure counts per step, revealing where to look but not what happened. The most critical signal, often skipped, is the record itself: canonical bytes of the result, hashed with SHA-256, serving as evidence rather than just a claim.

Proving the Run with AER-1 Receipts

Brennan Zambo, creator of Agent Execution Receipts (AER), argues that observability and receipts are two halves of one job. While dashboards watch the run, receipts prove it afterward to auditors or other agents. The AER-1 standard closes the vocabulary gap by allowing checkers to report five specific anchor outcomes: verified, verification failed, rejected: unresolvable record, not checked, and absent. 'Absent' is the honest term for a step that left no record, preventing systems from painting their own gaps green.

Case Study: The MAPPING.md Verification

On November 22, 2026, a doc_intel call analyzed the MAPPING.md file from an interop lab pull request. The system recorded the run at 2026-11-22T09:05:56Z, reading 9,237 bytes. The resulting analysis was minted as receipt 34d47e29-1b53-40a9-84df-c9cac9841245. Recomputing the SHA-256 hash over the 3,580 canonical bytes of the receipt yielded an identical match to the published output_hash, resulting in a 'verified' anchor outcome. This recompute is the critical line between observation and proof.

The Honest Limits of Observability

Observability describes a run but does not grade it. A perfectly observed execution can still be wrong, inefficient, or expensive. A receipt proves the saved result wasn't altered after hashing, but it never proves the result was correct. Furthermore, sampling pipelines that capture only one in ten calls have observability of just a tenth of the run. Provenance inside a record remains a producer assertion; the design goal is to build checks so robust that the assertion itself becomes unnecessary.

Key Takeaways

  • Chat transcripts are self-narrations, not observability; external records are required.
  • AER-1 receipts provide cryptographic proof via SHA-256 hashes of canonical bytes.
  • Observability must account for 'absent' records to avoid hiding gaps in data capture.
  • Verification is free and public for anyone, ensuring auditability outside the operator's team.

The Bottom Line

If your agent stack cannot produce a recomputable receipt for every step, you don't have observabilityβ€”you have anecdotes. Stop optimizing for live dashboards and start optimizing for the stranger who needs to verify your work tomorrow.