Skip to content
New guide: how to scope an AI project before you build

Intelligence

AI Agents

Agents that do real work across your tools: reading, deciding and acting inside limits you define, with every step observable.

Typical timeline
5-9 weeks
Group
Intelligence
First step
Fixed-price discovery

Scope, price and acceptance criteria are agreed in writing before the build starts.

Why us

Why choose us for this work?

A narrow job

We define one task with a clear start and finish, rather than an open-ended assistant.

Explicit permissions

The agent can only call the tools and change the records you allow.

Human in the loop

Anything consequential is proposed first and approved by a person.

Readable traces

Every step, tool call and decision is logged so a human can follow it.

Reversible actions

Writes are idempotent and can be rolled back when a step fails.

What you get

What you get from AI Agents

An agent with a job, a boundary and a trace.

  1. 01

    A working agent

    One end-to-end task running against real systems in a test environment.

  2. 02

    A permission model

    Tool access and data scope written down and enforced in code.

  3. 03

    An exception queue

    Cases the agent cannot resolve land with a person, with context attached.

  4. 04

    Observability

    Traces, costs and success rates visible per run.

  5. 05

    Handover notes

    How to change the job, the rules and the tools without us.

The problem

What usually goes wrong

An agent that can call any tool at any time is a liability. Useful agents have a narrow job, explicit permissions, a clear stopping condition and a trace a person can read when something goes wrong.

Who this is for. Teams with repetitive, rules-heavy work that currently depends on people copying information between systems.

Signs you need this

If any of this sounds familiar

  • The same task is performed by hand across three or more tools.
  • The rules exist, but only in somebody’s head.
  • Exceptions are common and need a person to decide.
  • You need a full record of what the system did and why.

How the work runs

How we deliver it

  1. 01

    Map the task

    We watch the work being done and write down the steps, rules and exceptions.

  2. 02

    Define the boundary

    Permissions, approval points and stopping conditions are agreed in writing.

  3. 03

    Build and trace

    The agent is built with a readable trace for every run.

  4. 04

    Stress the exceptions

    Conflicting data, failed tools and ambiguous instructions are tested deliberately.

  5. 05

    Release and hand over

    We launch behind a flag, watch the measures, then hand the controls to your team.

Deliverables

What you receive

  • Task map and decision rules
  • Agent implementation with tool permissions
  • Approval flow and exception queue
  • Tracing, logging and cost controls
  • Runbooks and handover documentation

Stack

Tools we expect to use

TypeScriptPythonLLM APIsEvent queuesPostgreSQL

We choose the stack to fit the problem. If your team already runs something else, we plan around it.

Questions

Common questions

Can we turn it off?

Yes. The agent runs behind a flag, and any write can be paused or reverted without changing the rest of the system.

What happens when it is unsure?

Uncertain or high-impact cases go to a person with the context attached. The agent proposes; it does not decide alone.

Will it replace our people?

It removes the copying and checking steps. The exceptions and judgements stay with your team, with better information.

Next step

Talk through your project.

A short conversation. You leave with a clear view of what it would take, and whether this service is the right first step.

Prefer email? Write what you are working on and we will reply.

Send a note

By sending this, you agree that we may reply to your enquiry. See our privacy policy.