Evaluating AI agents for production

What to put in the set

Real historical cases with known outcomes—including ugly scans, partial packages, and hostile edge cases.

Label field-level truth where extraction matters. End-to-end "looks good" is not enough for money fields.

Metrics that matter

Precision/recall on critical fields, hard-rule violation rate, human override rate, and time-to-decision—not only BLEU or vibe scores.

Regression gates in CI: config changes that drop critical-field accuracy do not ship.

Building the set

Pull real historical packages. Label critical fields. Include the ugly cases: bad scans, missing pages, adversarial free text.

Version the set. When policy changes, update labels and re-run gates before promoting config.

CI gates

A config or model change that drops precision on money or identity fields should fail the pipeline. Manual "looks fine" is not a gate.

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.