AI agents
Agents that do real work, within limits you set.
An agent reads, decides and acts across your tools. That is useful only when you know exactly what it may do, when a person steps in, and that it still works after every change. We design, build and prove all three.
Where agents earn their place
The jobs that suit an agent are repetitive, spread across several tools, and have a clear line where a person should take over.
Support and operations
Read the order, check the policy, draft the refund or the reply, and hand anything unusual to a person.
Documents and forms
Pull the facts out of contracts, invoices or applications and put them into the systems that need them.
Research and reporting
Gather what a weekly report needs from your tools, draft it, and flag what changed since last week.
Internal assistants
Answer staff questions from your own documents and take the small actions that follow, such as opening a ticket.
What is in an agent we build
Every agent ships with the same six parts, so that it can be trusted, checked and handed over.
Design sheet
What the agent may do, what needs a person to approve, and what it must never do. Agreed before we write code.
Tool connections
Logged access to your systems, with the narrowest permissions that do the job.
A test set
Realistic cases, edge cases and adversarial inputs, scored on every change.
Approvals and logs
A person approves the actions that matter, and every tool call and answer is recorded.
Monitoring after launch
We watch quality and cost once the agent is live, and re-test when a model or a tool changes.
Hand-over
Your team owns the code, the test set and the documentation, written so that they can change it.
How a build runs
The same four steps as everything we do, with a fixed price agreed after the first one.
01Design
A kickoff call, then a design sheet: the job, the users, the risks, what the AI may and may never do, and what a good result looks like. We agree the scope and a fixed price.
02Build
In your stack and your repository, with the test set written alongside the code. You see working software early, not at the end.
03Prove
Every change is scored against the test set, with safety checks in an agreed scope. You get a report and a go or no-go recommendation before launch.
04Run
After launch, every prompt, model or tool change is re-tested. You get a monthly report on quality, cost and what changed.
Questions we get asked
Which models and platforms do you use?
The ones that fit the job and your constraints: hosted models from the main providers, or open models you run yourself. We are not tied to a vendor.
Can the agent work inside our existing product or tools?
Yes. Most agents we design live inside a product or a workflow tool the team already uses, rather than in a separate app.
We already have an agent. Can you test it?
Yes. We can evaluate an agent your team built, or take over its upkeep. See AI evaluation & QA.
AI evaluation & QAHow is the price set?
It is fixed, and agreed after the design step, once the scope is written down. No hourly billing.
Have a job in mind for an agent?
Tell us what it should do and what it must never do. We will come back with questions, then a design proposal.
Founding clients get a lower price on their first project.