Production is graded, not demoed
Every AI agent looks good in a demo. The demo is a fixed scene: the vendor picked the input, rehearsed the flow and chose the moment. The question that separates a deployment from a toy is different: on any given Tuesday, with no vendor in the room, how do you know the agent still does its job correctly? The honest answer, for anything that runs a real workflow, is an evaluation harness: a test suite that grades the agent continuously, wired into the release path so nothing changes until it passes.
What an evaluation harness actually is
Four moving parts, none optional:
Golden cases. A curated set of real inputs and the correct outputs, drawn from actual workflow decisions rather than invented examples. For a quotation agent, that is last quarter’s real rate requests with the quotes your team actually approved.
Regression sets. Every bug ever caught becomes a case that fails if it ever returns. The set only grows, which is the point: it is the memory of everything that has gone wrong.
Quality gates in the release path. Eval results do not live in a dashboard someone might read. They block the release. A change ships when the evals pass, not when the demo looks good.
Drift measurement. The workflow the agent was built for changes, and so does the world it reads. Drift is measured against the day-zero baseline, so quality claims stay checkable months after go-live instead of resting on launch-day anecdotes.
Why the attention is moving here now
Coding agents went mainstream, and agent output is now code: pull requests, migrations, configuration. That forced the industry to put review and evaluation tooling into the delivery path, because unreviewed code from any author is a liability. The same logic applies to operational agents. An agent that quotes freight, processes applications or chases invoices is making consequential decisions many times an hour. It needs the same discipline code needed, with one difference: its test suite grades judgement, not just syntax.
What good looks like, in five checks
- The eval suite lives in your repository, versioned beside the code, not on the vendor’s laptop.
- Golden cases come from your real decisions, and you can read every one of them.
- Quality gates block releases mechanically, including the vendor’s own changes.
- Drift is reported against the measured baseline your board funded.
- Rollback is rehearsed, because a failing eval should be a routine Tuesday, not an incident.
Five questions to ask any AI vendor
Who owns the evaluation suite if we part ways? What happens when an eval fails at 2am? Where did the golden cases come from, and how many are there? How is drift measured against the baseline over time? Show me a release your own evals blocked. A vendor that has never blocked one is not evaluating anything.
The Nugawi angle
Nugawi, an applied-AI consultancy in London, ships the harness as part of every 90-day production build: golden cases from your real workflow, a regression set that starts before cutover, quality gates in the release path, and a final readout that re-measures the workflow against the day-zero baseline. If your agents are already live and the five questions above feel uncomfortable, that gap is exactly what the Agent Audit covers. If you are starting from zero, the 90-day production build is the path, evaluation harness included.