Abstract

Fix this bug. Answer this customer. Onboard this employee.

Each request has a beginning and, with some agreement about the result, an end. A person has already decided that the work matters and handed it to the agent.

Now change the request. Keep the support queue healthy. Make sure every new employee is ready for their first day. Maintain this service.

The agent has to notice work arriving, decide what needs attention, and recognize when something it previously completed needs another look. A customer replies with new information. An employee’s start date changes. An urgent request interrupts work already in progress. The agent’s original instructions may no longer describe the situation in front of it.

Coding gives us a useful comparison. An agent that fixes an issue can submit a patch and stop. An agent responsible for a repository has to deal with future regressions, competing priorities, and changes in what its maintainers want. The same distinction matters when we move agents into support, finance, and operations.

This talk examines what changes when we give an agent continuing responsibility. We’ll explore how to define the condition it should maintain, respond to events, preserve progress across interruptions, and hand work to a person without losing context. We’ll also examine how to bound its time, spending, and authority, and how to revise or withdraw a delegation that is no longer appropriate.

Attendees will leave with a framework for deciding which responsibilities an agent can sustain, how to supervise them, and how to evaluate its performance over time. Completing yesterday’s tasks is useful evidence. It does not tell us whether today’s queue is under control.

Key Takeaways

• Distinguish a request with a defined endpoint from a continuing responsibility that requires attention over time.
• Define goals around the condition an agent should maintain, including acceptable delays and unresolved exceptions.
• Design for new events, changing priorities, interruptions, and reliable handoffs between agents and people.
• Establish limits on spending and authority, with explicit conditions for pausing, escalating, or ending a delegation.
• Evaluate sustained performance using missed work, rework, exception age, and human intervention alongside completed tasks.

Questions

• What changes when an agent must decide that work needs doing before anyone asks it?
• How should an agent respond when new information invalidates a task it already completed?
• What should happen to work in progress when a higher-priority request arrives?
• How can a person take over a responsibility without reconstructing the agent’s entire history?
• What evidence would show that an agent is keeping a queue healthy rather than accumulating unresolved problems?