AI agent development for one job, done unattended
I build an AI agent that does one defined job inside your business, with the checks, the logs and the handover a person would need to trust it running without them.
- You bring
- A job a person does today, and the systems it touches
- You get back
- An agent in production, with its checks, logs and runbook
- How it is cut
- Three milestones: one path working, the full job, the handover
What you get
- The agent itself, in your repository, running on your infrastructure or mine until handover.
- Tool definitions for every system it touches, each with a written statement of what it may and may not do.
- A gate before any model call for input that should never reach a model, and a gate after it for output that should never reach a customer.
- A record of every run: what came in, what the model was asked, what it answered, what the agent did, attributed and append-only.
- An automated test suite that runs on every push, and a runbook for the day the provider is down.
How the work is done
Start from the job
The first document I write back to you describes the job as a person does it today: the inputs, the decisions, the exceptions, who is told when it goes wrong. The model is chosen after that, as a configuration decision that can change.
One path first
The first milestone is the commonest path through the job, working end to end on your real data. It is small on purpose: you can run it, and we both learn where the model is weak before the rest is built around that.
Decide where a person stays
Some steps a model does every time and some it fails one time in two. I measure which is which on your task and put a person, a rule or a refusal at the weak steps, so where the system needs a human is a design decision with numbers under it.
Write down what it did
An agent that acts for you has to be able to answer for it later. Every run is recorded so a wrong outcome is a lookup, and so the system can be evaluated at all. This is the subject of my paper at the HAIC workshop, NeurIPS 2026.
Where I have done this
The 1o1.org production system: a dialogue engine with safety checks before any model call and about 370 automated tests, built and run by me since April 2026.
An unattended video engine that found a subject, wrote, rendered and published to three platforms, running five accounts daily with nobody approving each step.
36 models graded on 20 office tasks, 720 runs: the measurement of which tasks a model can be left alone with.
AIGROW benchmark, 23 Sep 2026(verifiable source, opens in a new tab)
Who it fits
A good fit when
- The job is real and someone does it today, so there is a definition of done.
- You can give the agent access to the systems the job touches, or a sandbox of them.
- A wrong action has a cost you can name, so we know which steps need a gate.
Not a fit when
- You want a chatbot on the website and the job is "answer anything".
- Nobody can say what a correct outcome looks like yet.
- The work needs someone sitting in your team day to day.
Questions about AI agents
- What does an AI agent cost to build?
- It depends on how many systems the agent touches and how expensive a wrong action is, because those two decide how many tools and gates have to be built and tested. I quote per project, in writing, after reading your brief, with a price per milestone. You approve each milestone before the next one starts.
- Which model do you build on?
- Whichever one passes your task at the lowest cost, which I measure before committing. The agent talks to the model through one interface, so changing provider later is a configuration change and a rerun of the tests. My production systems reach their models through OpenRouter, which keeps Claude, GPT and open models one configuration line apart.
- How do I know the agent will not do something it should not?
- Each tool it can call has a written scope, input that should never reach a model is stopped before the call, and every run is recorded. On top of that, the steps where the model fails often on your task get a person or a rule in front of them. The model is no more reliable for it, but the system around it can now account for every action.
- Can you take over an agent somebody else started?
- Yes, and the first milestone is then a written review of what it does today: what it records, what it can overwrite, where it fails. Often the model is fine and the write path is the problem.