---
title: "How we build an AI agent, in five steps"
updated: "2026-11-17"
url: "https://www.etheon.com/research/index/how-we-build-an-ai-agent"
authors: "Etheon"
published: "2026-11-17"
type: "research"
---

# How we build an AI agent, in five steps

An agent is software that takes actions in your systems. We build it in five steps, inside limits held outside the model, with a person where it matters.

An agent is software that takes actions in your systems: it reads a task, calls the tools it is allowed to call and records every step. We build each agent in five steps, inside a harness of permissions and limits that sit outside the model, with a person approving the steps you choose, so the agent can only do what you have set.

## What are the five steps?

1. Scope the task: the steps, the systems and the decisions a person keeps.
2. Build the tools: each tool the agent may call, with its own permission.
3. Evaluate: an evaluation set from your own cases, scored before release.
4. Release in stages: first a person approves every action, then only the exceptions.
5. Watch and improve: traces, alerts and a regular review of what it got wrong.

## What limits does the agent work within?

The limits you set, enforced outside the model:

- Tools the agent may call, each with its own permission.
- Limits on data, spend and scope.
- A person who approves the steps you choose.
- Every step logged, so your team can replay what the agent did and why.

The software around the model that holds those limits is the agent harness: the tools it may call, the permissions on each, the context it may read, where it runs, the steps a person approves and the log of every step. We build the harness first and the agent inside it, so its limits hold whatever the model answers. Our research note [The harness comes first](https://www.etheon.com/research/index/the-harness-comes-first) sets out why.

## What does a run look like?

Every step is written down. Here is the format, on an illustration (not a client's system): a refund that the agent may make only within its limit.

| Time | Step | What happened | Result |
| --- | --- | --- | --- |
| 00\.00 | request | Refund order 1042 if the policy allows it |  |
| 00\.12 | retrieve | refund-policy.md, section 3.2 | grounded |
| 00\.31 | tool | orders.lookup(1042) | delivered 9 days ago |
| 00\.44 | check | inside the 14-day window | pass |
| 00\.52 | guard | amount under the agent's limit | pass |
| 00\.53 | person | approval not needed under the limit | skipped |
| 00\.71 | tool | payments.refund(1042) | done |
| 00\.72 | log | every step saved for review | saved |

## How is the agent tested before release?

On an evaluation set built from your own cases, scored before release. Like every model we build, it passes the same gates before it reaches your users: task quality, grounding, safety, privacy, latency, cost and drift. We agree each threshold with you before work starts.

## What do you receive?

- The agent and its tools in your repository.
- A permission map.
- An evaluation set and its scores.
- A trace of every run.
- A runbook for your team.

## Where does an agent we built run today?

Jeriko is client work: Etheon built it for a client, and we name it with the client's permission. In its own words, Jeriko is a team of personal AI agents, each a person in the app with its own memory, instructions, wallet and email address. On our AI page it is the proof of agents that act within the permissions you set. Its story is on [Work](https://www.etheon.com/work/jeriko).

The agents stage, the harness and the release gates are on our [AI page](https://www.etheon.com/services/ai#agents). Agents for your company: tell us the task and the systems it touches. We reply within one business day.