An AI agent is a program that uses a language model to decide its next step: which tool to call, which record to read, when to answer and when to pass the case to a person. A company needs one when a process has too many variations for fixed rules and too much volume for people alone, and when a wrong step has a cost that someone can name.
What I build
- Agents for one process at a time. Admissions questions, quotations, support cases, internal requests. One agent, one process, a clear list of what it may do.
- Multi-agent systems. Several agents with separate tasks behind one conversation, with routing, shared state and handoffs between them.
- The tools the agent uses. Connections to your databases, CRM and internal systems, each with its own permissions, built on the Model Context Protocol where it fits.
- The evaluation that decides a release. Cases written as code, deterministic checks, a judge model calibrated against people, and a gate that blocks a release when the agent gets worse.
- Operation. Deployment, logs of every tool call, and a way to see what the agent did and why.
Signs that a process is ready for an agent
- The work arrives in language (messages, documents, requests) and ends in an action in a system.
- People already follow a procedure for it, even an unwritten one, and can tell a good outcome from a bad one.
- The data and the systems it needs can be reached through an interface, with an account that can be limited.
- A wrong answer can be caught: by a rule, by a person or by a later step.
When one of these is missing, that piece is the first work and the agent comes after it. The order matters more than the model: a company that describes its process first gets an agent it can check, and one that starts from the model gets a demonstration.
How the work runs
- The process first. I read how the work is done today, where it breaks and what a correct outcome is. An agent placed on a process nobody has described repeats its confusion faster.
- Cases before code. We write the situations the agent must handle and the outcomes that are not acceptable. These cases become the tests.
- A first agent on real data. A narrow version runs against the cases, with the real tools and a copy of the real state.
- The gate. Before any release the suite runs. A failure in a required case stops the release, whoever asks for it.
- Release and watch. The agent goes to a small share of the work first. What it does is recorded, and new failures become new cases.
What you receive
The agent and its tools in your own repositories and your own cloud account, the cases and the evaluation harness, the deployment, and written notes on how to operate and change it. Nothing depends on an account of mine.
What this does not include
I do not promise that an agent replaces a team, and I do not put an agent in front of customers without tests that say what it may not do. If the process is not ready for an agent, I say so, and the first piece of work is the process.
Where the evidence is
The agents of a higher-education group that serves 700,000+ students run in production on Vertex AI, the Agent Development Kit and Cloud Run. My part was the evaluation harness, the calibrated judge, the isolated state for every case and the gates of each environment. The case study and the guides linked below describe it, with what it shows and what it does not.