AI agent observability

Minimum viable logs

Input identity (case id, document ids), model and prompt/config versions, tool calls with args redacted as needed, check results, human actions, final state.

Store structured events, not only chat transcripts. Transcripts help debugging conversation; they do not prove a control ran.

Traces vs evals

Traces tell you what happened on one run. Evals tell you whether the system is still fit for purpose on a held-out set of real cases.

Both belong in the operating model. Shipping without either is how silent regressions land in production.

What to alert on

Spike in override rate, drop in critical-field precision, growth in dead-letter depth, and tool-call error bursts. Model latency alone is not enough.

Alert the workflow owner, not only the platform channel. Ops must see the queue, not a generic p99 graph.

Privacy in traces

Redact secrets and apply field-level policy for PII before traces leave the trust boundary. Observability that violates retention rules will be turned off—and then you are blind.

In practice

Map the workflow on a whiteboard before you open a framework: inputs, systems of record, humans, and irreversible writes. If that map is fuzzy, the agent will encode the fuzz.

Pick ten to fifty real historical cases as an eval set. Include the ugly ones. Run the agent offline against them until critical fields and hard rules are acceptable. Only then connect write tools.

Ship with a pause switch, a human queue, and a weekly review of override reasons. Promote repeated overrides into rules. That loop is how production systems improve—not another prompt brainstorm.

Common failure modes

• Treating a demo on clean samples as readiness for production volume.
• One shared service account with broad write access across systems.
• No owner for the exception queue, so failures pile up as noise.
• Changing prompts and models without regression gates on real cases.
• Measuring only model latency or thumbs-up, not completed-case cost and audit completeness.

What good looks like after ninety days

The first workflow is boring: stable override rate, known failure modes, operators who trust the queue. Config changes go through review. Traces answer "what happened to this case?" without archaeology.

At that point you can add a second document type or a second agent role. Expanding before the first path is boring is how programs stall with five half-built pilots.