Designing for Agents: When Software Starts to Act

For most of the history of software, systems waited. They waited for a person to click, to type, to approve. Intelligence, where it existed, was advisory: the system suggested and the human acted.

Agentic AI changes that contract. An agent pursues a goal through a sequence of actions — calling tools, gathering information, making intermediate decisions — without a person directing every step. That is a genuine shift, and it deserves more careful design thinking than the current excitement usually allows.

Permissions before intelligence

The most useful early question about an agent is not about the model. It is about authority. What is this agent permitted to do? Which systems can it touch, and with what scope? What must it never do without a person’s confirmation?

Writing this down — a short operating contract for the agent — has a clarifying effect. It forces the team to distinguish between actions that are cheap to reverse and actions that are not. Sending a draft for review is recoverable. Sending money, deleting records or making a promise to a customer is a different category, and the design should treat it as one.

Evidence and recovery

When a person completes a task, we rarely ask them to prove it. When an agent completes a task, we should. Evidence of completion — what was done, on what basis, with which sources — is what makes delegation to software auditable rather than hopeful.

Just as important is the recovery path. Agents will encounter situations they cannot resolve: ambiguous instructions, missing data, systems that fail mid-task. A well-designed agent knows how to stop, how to say so, and how to hand back control with enough context for a person to continue. The measure of a mature agentic system is not that it never fails, but that its failures are visible and recoverable.

Where agents belong — and where they do not

There is a temptation to make everything agentic. Resist it. Processes that are already well understood are often better served by explicit, predictable workflows: cheaper to run, easier to audit, boring in the best sense.

Agents earn their place where flexibility creates value — where the path to the goal varies, where judgment is needed at intermediate steps, where a rigid process would break more often than it would help. The organisations that get this right will not be the ones with the most agents. They will be the ones that chose deliberately where autonomy belongs.