AI agents & systems

A possible starting point is a research assistant that gathers material from agreed sources and prepares a draft with references for review. Another is an internal assistant that uses project context to prepare a handover or propose the next step. We could explore either approach using representative tasks from your work.

Before selecting an architecture, we define the task, available information and acceptable output. A fixed workflow may already be enough. An agent becomes useful when the work calls for choosing between steps, using tools or responding to new information.

We work out which information the system needs at the point of use, which tools it may access, and what should persist between tasks. Memory needs a purpose: retaining relevant project context is different from collecting everything indefinitely.

Bring representative tasks, example outputs and an outline of the systems involved. We discuss permissions, data boundaries and the person responsible for reviewing results. The first deliverable might be an architecture recommendation, a working prototype, or a build with an agreed scope and budget.

An agent needs an explicit response to missing evidence, unavailable tools and requests outside its scope. We identify actions that require approval and make room for people to correct or stop the work. Adding autonomy should follow a demonstrated need.

For a build, we test against representative tasks and examine the failures as well as the useful results. Handover covers operation and the agreed limits, so you know what the system is intended to do and where human judgement remains part of the process.