AI agents
Systems that take a goal, use your tools, and complete multi-step work with a human checkpoint.
- Engagement
- Phased build, then a supervision retainer
- Typical timeline
- 6–8 weeks
- Starts with
- A list of the actions the agent will be allowed to take, and the ones it never will.
An agent is software you hand a goal to, not a task list. That is the appeal, and it is also the risk: an unsupervised one can spend real money, email a real customer or overwrite a real record before anyone notices it misread the situation.
Work that finishes overnight without a person driving it.
What we build
Concrete artefacts, handed over and documented.
A written list of every tool the agent may call, and the limit on each, which records it can edit, which amounts it can commit, which people it can contact
Approval gates on anything you would not want reversed, with the request shown in plain language to the person approving it
A run ledger holding every step, tool call, input and output, so any outcome can be reconstructed months later
Stopping rules, a step budget, a spend ceiling and a timeout, so a confused agent fails loudly instead of looping
A replay harness that reruns past cases against a changed prompt before the change reaches live work
Cost per completed run tracked as a metric rather than discovered on an invoice
What changes
Multi-step work finishes without a person holding it open
Every action carries an audit line naming the agent, the input and the approver
The worst financial case for a bad day is bounded and known in advance
How it runs
- Week 1
01Scope and permissions
We agree the goal, the tools and the exact boundary between what runs freely and what waits for a human.
- Weeks 2–5
02Build and constrain
The agent, its tools and its guardrails, tested against real cases including the ones designed to confuse it.
- Weeks 6–8
03Supervised run
It works alongside your team with every action reviewed, and the approval gates loosen only where the record supports it.
Chosen per project. Named here so you can see the shape of it.
- LangGraph
- Anthropic Claude
- Temporal
- Postgres
- Langfuse
Questions we get asked
Tell us the outcome, not the tooling.
Send us the situation you are in. We will tell you which discipline it belongs to, what we would do first and what it costs, including when the answer is to wait.