---
title: "AI Harness Development Week by Week: What a Real Project Looks Like, From Diagnosis to Production"
description: "AI harness development week by week: what happens in each phase, what you receive and which decisions are yours on the way to AI agents in production."
url: https://addabra.com/en/blog/ai-harness-development-week-by-week
lang: en
date: 2026-11-01
category: "Services"
keyword: "AI harness development"
publisher: Addabra
alternate: https://addabra.com/blog/desarrollo-de-ai-harness-semana-a-semana
---

# AI Harness Development Week by Week: What a Real Project Looks Like, From Diagnosis to Production

> AI harness development week by week: what happens in each phase, what you receive and which decisions are yours on the way to AI agents in production.

Monday, 9:00 a.m. The COO of a professional services firm opens her inbox and finds the third proposal this quarter to "implement AI agents." All three promise results. None of them explains what will happen in week two, what her team needs to do, or how she will know halfway through whether the project is on track.

That uncertainty is a big reason so many projects stall at the pilot stage. This article walks through what **AI harness development** looks like week by week: what gets built in each phase, what you receive at the end of each one, and which decisions are yours.

## Before you start: what exactly gets built

An AI agent is a model plus a harness. The model reasons; the **harness** is everything around it that makes it work reliably: what information it sees, which tools it can use, which rules it cannot break, who approves its decisions and how its actions are logged. For the full map of components, see [anatomy of an enterprise harness](https://addabra.com/en/blog/anatomy-of-an-enterprise-harness).

An AI harness development project does not build "a chatbot." It builds the foundation every one of your agents will run on, and it launches that foundation with a first real flow. That difference shapes the whole timeline.

## The timeline, phase by phase

What follows is an **illustrative timeline**. Actual duration depends on the process, your systems and the quality of your data. What does not change is the order, and that every phase ends with something you can verify.

| Phase | Weeks (illustrative) | What you receive |
|---|---|---|
| Diagnosis | 1–2 | Prioritized process, systems map and success criteria |
| Design | 3 | Architecture for agents, memory, rules, approvals and evals |
| Build | 4–7 | Harness and first flow running on real cases |
| Production | 8 | Real users, live approvals, traces and alerts |
| Evolution | Ongoing | Data-driven tuning and new agents on the same foundation |

### Weeks 1 and 2: diagnosis

No agent code gets written here. The team interviews the people who do the work today, reviews the systems they touch (CRM, ERP, email, documents) and identifies the exceptions: the odd cases a demo never shows.

The outcome is a decision: **which process goes first**. The best candidate usually has volume, clear rules and a visible cost when it is done late or done wrong. We go deeper on how to choose it in [process diagnosis for AI](https://addabra.com/en/blog/process-diagnosis-for-ai).

Your part: grant access to the people and the systems, and name a process owner with the authority to decide.

### Week 3: design

With the process chosen, the system is designed on paper before it is built:

- **Agents:** which specialized agent handles each part of the work.
- **Memory:** what the business knows about each client or case, and where that information lives.
- **Rules:** the list of "nevers" the system must enforce.
- **Approvals by role:** at which steps a person signs off, and which person.
- **Evals:** the set of real cases used to measure whether the agent does its job well.

This is the moment to argue about the hard cases. Changing a rule in a document costs a conversation; changing it once the system is live costs far more.

### Weeks 4 through 7: build

The build moves in short cycles with a weekly review. Every week your team sees the flow running on real cases, not slides. A typical sequence:

1. The memory and the connections to your systems are set up.
2. The first agent is built and measured against the evals.
3. The remaining agents in the flow are added, along with the orchestration between them, including retries and recovery from the last completed step.
4. Approvals by role, traces and spend caps are switched on.

Your part: attend the weekly review, correct criteria and bring hard cases. Evals get much better when the team brings the examples that caused them real pain.

### Week 8: production

Go-live is gradual: first a group of users, a limited volume and every approval switched on. Traces are reviewed daily, cost per flow is tracked and performance is compared against the evals agreed on during design.

A good sign the project is healthy: when something fails, the team can see which agent did it, with which data and at which step, without having to ask anyone.

### After that: evolution

The first flow is the harness's debut, not the end of the project. With the foundation in place, the second agent reuses the memory, approvals, evals and observability. That is why each new agent should cost less and behave consistently with the others.

## Illustrative example: the same project, with and without a method

Imagine a distribution company that wants to automate responses to requests for quotes.

**Without a method:** a team builds a pilot in two weeks using a powerful model and a pricing spreadsheet. In the demo, it is impressive. In operations, it quotes outdated prices, has no idea which clients have special terms, and nobody reviews anything before it goes out. Two months later, the pilot is switched off.

**With a method:** the diagnosis reveals that the real problem is pricing exceptions. The design puts those rules into the harness and requires the VP of Sales to sign off on any off-list quote. The build measures each version against real cases from the last quarter. In production, the agent prepares and a human decides. We explain the pattern behind this difference in [why your AI pilot never reached production](https://addabra.com/en/blog/ai-pilot-to-production).

## Common mistakes in an AI harness development project

1. **Skipping diagnosis.** Building before choosing the right process guarantees a good-looking, low-value pilot.
2. **Leaving approvals for the end.** They are part of the design, not a safety patch.
3. **Made-up evals.** If the test cases do not come from your real operations, they measure a business that does not exist.
4. **Reviews made of slides.** If you do not see the flow running every week, you do not know whether the project is moving.
5. **One harness per agent.** Rebuilding the foundation for every use case multiplies cost and contradictory rules.

## What to do on Monday

Pick a candidate process and write down three things on a single page: how much volume it has, which exceptions complicate it and who should sign off before anything reaches a client. With that, any conversation about agents starts in the right place.

At Addabra we work with exactly this method: diagnosis, design, build, production and evolution, on our own proven harness and with visible deliverables in every phase. To see what it includes, take a look at our [AI harness development service](https://addabra.com/ai-harness-development/en).

## Frequently asked questions

### How long does an AI harness development project take?

It depends on the process you pick, the state of your data and how many systems need to be connected. The timeline in this article is illustrative; what matters is that every phase ends with a concrete deliverable and that the first flow reaches real users as early as possible, not at the end of a long plan.

### How much time does my team need to commit?

Mostly the time of the people who know the process: interviews during diagnosis, a rules review during design and a weekly review during the build. You also need an owner with the authority to decide what gets approved and who signs off on each step.

### What happens after go-live?

The evolution phase begins: the flow is tuned with operating data, the evals are updated and new agents are added on the same foundation. Each new agent reuses the memory, approvals and observability that already exist.

## About Addabra

Addabra is an AI-software-first software factory that designs and builds custom agentic systems for companies: one agent per process, a shared business memory, orchestration, human approval by role, auditing and cost control. Free diagnosis: https://addabra.com/en/blog/ai-harness-development-week-by-week#contacto