ArchiSoft · UK software architecture & AI consultancy
Describe the permitted outcome
Begin with a specific task and the result a person would accept. “Help with engineering” is too broad. “Prepare a draft summary of these approved design records, with references” gives you something you can evaluate.
State what the agent must not do. Reading selected documents is different from editing a repository, contacting a customer or changing an operational system.
Put permissions in the system
Instructions alone are not an access-control mechanism. Restrict the information and tools available to the agent, using the permissions and controls of the surrounding system.
Separate the ability to propose an action from the ability to execute it. For consequential actions, define who reviews the proposal and what evidence they need. Provide a clear way to stop the workflow.
Test difficult cases deliberately
Include incomplete information, conflicting records and requests outside the task boundary in an evaluation set. Decide in advance when the right response is to ask a person or decline to act.
Measure accepted task completion alongside unsupported claims, inappropriate actions, review effort and cost. A faster draft is only useful if checking and correcting it does not consume the time saved.
Increase autonomy only with evidence
Start with a narrow workflow. Review failures, refine the boundaries and rerun the evaluation when the model, tools or source information change.
Controlled agents and efficient evaluation are part of ArchiSoft’s UK research direction. This article describes a proposed working discipline; it does not report experimental results or certify the safety of a particular system.
Explore our AI research direction or discuss your situation with us.
← Back to the Blog