Skip to content
Systems that act

More than an answer. Still within a boundary.

An agentic system can take several steps towards an assigned task. The important question is not how autonomous it sounds, but what it may actually do.

Research note

Break the task into visible steps.

Consider preparing a website improvement: collect permitted information, identify a problem, propose a change and ask for review. Those are distinct steps with different failure points.

Research note

Give each step its own permission.

An assistant allowed to read a page should not automatically be allowed to publish one. Set the accessible project, permitted tools and stopping conditions before it begins.

Research note

Design the stop as carefully as the start.

If a source is missing, the next action is unclear or a tool fails, the system needs to stop or ask. Retrying a payment or publication without knowing the previous result can cause harm.

Research note

Keep the explanation available.

Record which step ran, what it used and whether it succeeded. The person responsible needs a readable account of the work, not just a finished-looking screen.