Agents

Agents that finish a job a person can check.

An agent, in our work, is software that carries a task through to a result. It has a boundary. It can show what it read and what it wrote.

If it cannot show its work, it does not go into production.

Kap is the public example. A support worker opens the incident form their organisation already uses and talks through it in the hallway, while they still remember the detail. Kap asks for the empty required fields. The answers land in those fields. A coordinator can read them before anyone treats the record as done.

The same pattern applies outside that form. Define the job. Give the agent the tools it may call, and the documents it may read. Keep a trail of the run: what it retrieved, what it changed, what a person accepted. Voice belongs in this, because a lot of the work happens while someone’s hands are busy. The agent still has to come back with the structured result the system expects.

A chat window is a demo. We build the agent when the organisation can name the task, the permission, and the person who reviews the output. Retrieval sits underneath that, which is why the writing on this site is about RAG. An agent that answers from memory will sound finished and be wrong about your files.

Notes on retrieval

Talk to us about an agent