17 September 2026 · 7 min
The demo is not the system
Vendors will ship a new model name next month. If your process is 'paste into the latest chat', you do not have an agent. You have a habit. A production agent has: a fixed tool list, retrieval over your corpus, an evaluation set, logging, and a human who can stop it. The model is a dependency you pin, test, and swap.
How we keep it on a leash
Version the prompt. Version the model. Run the same 50 tasks after every change. If accuracy drops, do not ship. Keep PII out of the prompt unless the tenancy is yours. Do not let the agent write to gold. Let it draft, classify, extract - then a job with tests commits.
Public-sector teams ask for this because a GDS-shaped service cannot surprise a user. Private teams ask because a wrong invoice classification is money. Same harness. We have stopped more 'agent' projects at discovery than we have put in production, and that is the service.
Related service: AI & machine learning