An agent you can’t stop or check is a liability, whatever it produces.
Putting an agent into production means deciding what it may do, how its work will be checked and who can stop or recover it when something goes wrong. We build those controls with your engineers, on your infrastructure, so responsibility stays with people.
What our own system teaches us
We run our consulting practice with AI agents, and we can walk you through a sanitised account of its architecture, controls and the failures we have met, without exposing client data.
The main lesson: written instructions set expectations; they do not make an agent obey them. Permissions and checks outside the model constrain particular actions. Verification tests whether the work meets an agreed requirement. Human review covers decisions that need judgement. Each control has a scope and can miss failures, and an agent’s claim of completion is not evidence that its work is correct.
What you get
An agreed implementation with explicit acceptance criteria: permitted actions, evidence of changes, verification of results, and named human responsibility. We test failure and recovery as well as successful runs, and hand over what your team needs to maintain the system. Those checks are evidence about the cases tested, not a guarantee about every case.
How we work with your team
- Embedded: we build alongside your engineers, establishing permissions, validation and review in the working system.
- Advisory: your team builds; we review the architecture, decision boundaries and failure cases.
- Ongoing: a capacity retainer supports your team as the systems evolve.
What it costs
Discrete pieces of work are fixed price for a fixed timebox, agreed before work starts. Ongoing support is a capacity retainer, not billed hours. Published rates, expenses at actuals, and if an engagement fails to meet the agreed standard, we refund the fee.
Where to start
If the problem still needs investigation, start with AI Opportunity Discovery. If you know what you need to build, tell us about it, including what the agent would be allowed to change and what failure would cost.