Workflow Wednesday: Defining the Playing Field for AI Agents

Art by @basilonmypizza: https://lnkd.in/eF8FkWzN - https://basilhefti.ch/

As a kid playing soccer, this was the big disappointment: you scored a goal, only for the referee to call offside. You thought you were behind the invisible line … but apparently not. This maps well to our third TODO when working with autonomous agentic systems (remember, the first was about circuit breakers and the second about fallbacks): guardrails and monitoring.

Like the white lines on a soccer field, you need to define the playing field and the rules. The agent operates within clearly defined boundaries: what it is allowed to do, and under which constraints. Example: if the agentic system automates access permission requests, access to non-critical assets within the same department may be granted by the agent; for other requests, humans need to review. If the system supports city trip planning: ok to make proposals, but not to book.

And just like in a team, actions need to be attributable. You know who made the pass, who took the shot, who gets the yellow card. Same for agentic systems: every action must be linked, verifiable, and traceable.

Having rules is good. It is the first step. You also need the referee.

In other words: monitor what the agents are doing, and intervene when necessary. This is a process view: the monitoring includes the sequence of actions leading to the goal. This can become quite complex. Consider an agent supporting customer service: it is allowed to draft responses, but not to issue refunds. Under pressure to resolve a case, it might still trigger a refund workflow because it optimises for resolution.

The referee monitors and can stop the game. Intervene. Disallow goals (often to someone’s frustration, as seen in the agonising over every critical moment after the match).

Monitoring and guardrails are required: like supervision of humans, supervision of autonomous automatic systems is needed to prevent unwanted outcomes.

What’s the one match you wouldn’t miss this season (any Thun fans out there)?

• https://lnkd.in/d7Yemfyx, N. Tomašev et al, Intelligent AI Delegation, Deepmind, 2026

• Art: https://lnkd.in/eF8FkWzN https://basilhefti.ch/