Agents running against production systems

The situation

Clients kept coming to us with the same shape of problem. A process where every step needs judgement: read this document, check these three systems, decide what to do, write the result back. No single step is mechanical enough for a conventional workflow engine. The traditional approach ends up as a hundred nested conditionals that work on the first few examples and break on the fifth. The question we kept being asked: can an agent handle this end to end?

What we did

We treat agentic systems as production software. The agent plans before it acts, so improvising in production doesn't turn into a thousand-dollar API bill on a single task. The agent uses real tools against the actual business systems, with the right permissions, not a sandbox.

The search layer is tuned to the client's actual documents, because settings that work on one set never survive the next. Every agent ships with a labelled test set, so a bad prompt change is caught before deploy, not after. Every action gets logged as a decision record: what the agent saw, what it chose, why, what it changed.

Cost and rate limits sit at the orchestration layer, so a runaway loop stops itself. Humans review where the cost of a wrong answer is high, and come out where it isn't.

The shape of it

Delivered as pods and as embedded specialists inside client engineering teams.

The outcome

Agents running against production systems, with the audit trail to explain every decision and a labelled test set to prove a change hasn't broken anything.