What is an AI agent?
The architecture practice · Checked 12 Oct 2026
An AI agent is software that works one job in a loop. It observes the case, reasons about it, acts with the tools it's allowed, and verifies the result, until the job is done or it hands the case to a person. What keeps it safe is a written list of what it may reach, and one gateway that checks and records every action.
We call software an agent only when it does real work in a queue your team already owns. If it only answers questions, it's a report or a search tool, and we'll build that instead.
The loop: observe, reason, act, verify
- Observe. The agent reads the case and the material behind it.
- Reason. It prepares a decision with the tools on its list.
- Act. The gateway checks each call and writes it to the record.
- Verify. A reviewer approves before any change is committed.
fig 1 · agent-loopYour team writes down what the agent may reach
Every agent has a list of what it may read, what it may draft, and what it may never touch. Anything not on the list doesn't exist for it. The limits live in the gateway, so the agent can't talk its way past them.
agent: intake-agent acts_for: claims-team # a named owner may: - read: [policy, claim, party] - draft: [claim] may_not: - write: [payment] # a person signs escalate: above the limit you set -> reviewer record: every call, with its source
Six building blocks make up every agent we deploy
| Block | What it does |
|---|---|
| Where your team reviews the work | A console built for the role, or the collaboration tool you already use. |
| What your business terms mean | Only the terms the job touches, and which system holds the true definition. |
| One gateway for every action | It decides what each agent may do, records every action, and can serve read-only data. |
| Where the agent runs | A managed service, an instance destroyed after the job, or a person's workstation. |
| Every reviewer correction, kept | Each correction becomes a signal you can measure, not rework nobody sees. |
| Measurement from the first release | Correction rate, escalation rate, and the job's business number, against a baseline taken first. |
What an agent is not
| It looks similar | The difference |
|---|---|
| A general assistant for everyone | We turn this down. An agent works one queue that already has an owner. |
| A report | A report shows what happened. If a report answers the question, we build the report. |
| A job that runs twice a month | It doesn't need an agent. We'll tell you when a job doesn't. |
A reviewer stays until the corrections say otherwise
- Reviewed. In the first release, a reviewer approves every case, and we keep each correction.
- Promoted. Autonomy grows when the correction rate stays under a figure you set, for as many weeks as you set.
- Delegated. The agent works inside a written limit. A reviewer checks a sample and every exception.
All three stages are the same build, switched at the gateway.
Questions reviewers ask
Can an agent write to our system of record?
No agent writes inside your system of record. Every action goes through the gateway, and a reviewer approves before a change is committed.
Where does it run?
Where your data must stay: a managed service, an instance destroyed after the job, or inside your own cloud.
Can a regulator rebuild a decision?
Yes. Every case keeps its reasoning, and every action has a record from the first release.
How do we know it works?
We take a baseline first, then track the correction rate, the escalation rate, and the job's business number from the first release.