I Am the Agent's Guide Dog

The Agent is not my guide dog. It is more like a blind person with extraordinary abilities, moving quickly but unable to see the real world. I am its guide dog.

It can read a codebase in minutes, search thousands of sources, write software, operate a browser, and execute dozens of steps without tiring. Yet it cannot see the organizational history behind “make this feature simpler,” who will maintain the code, or whether users call a product difficult because it has too many buttons or because they distrust it. It can analyze only the reality that enters its input, and it need not live with the consequences. The danger lies in powerful action combined with limited perception of reality.

It Can Walk Perfectly in the Wrong Direction

Traditional software executes predefined instructions. Give an Agent a goal and it can find a route, use tools, recover from errors, and continue. That is progress, but an Agent may also proceed without knowing what it does not know.

It can execute a requirement that should never have existed, rapidly converting a careless sentence into code, pages, emails, and decisions. Every step may look reasonable while the journey ends at the wrong destination.

That is what suggests a guide dog. The dog does not know why its handler is going to a bank or station, and does not choose the destination. It can, however, see steps, cars, and overhead obstacles that its handler cannot. When an apparently valid command is unsafe, it may refuse. Guide-dog training calls this intelligent disobedience: refusing a dangerous local instruction to preserve the higher-level intent.

Holding the Harness Does Not Mean Seeing the Road

Handler and dog exchange information through the harness: the human gives direction, the dog reports the road; the human chooses the destination, the dog judges whether the immediate route is safe. This resembles collaboration more closely than “human controls machine.”

I should not remotely direct every file, tool, and line of code. My task is to state the destination—what should change, what success means, and which costs are unacceptable—and supply reality the Agent cannot see. A redundant process may carry organizational trust. An elegant design may leave maintenance to people excluded from the decision. “One more option” may signal distrust rather than a need for choice.

This knowledge comes from position, not superior intelligence. I inhabit concrete relationships, responsibilities, and consequences. Humans misjudge too, but in many tasks we remain closer to the actual street.

Humans Need to Learn Intelligent Disobedience

Much AI discussion asks Agents to refuse humans. Reuth Mirsky’s 2025 paper “Artificial Intelligent Disobedience” uses guide dogs to argue that a reliable AI teammate should seek clarification, offer alternatives, and refuse when commands conflict with safety, ethics, or the long-term mission. The metaphor also runs the other way: humans must intelligently disobey Agents.

When an Agent identifies the most efficient plan, I must ask whether its metric matters. When it produces a coherent decision, whose circumstances never entered the context? When cheap execution lets it keep expanding the task, has the problem already been solved? At critical junctions, I must check the destination, not merely whether it hit a wall.

Intelligent disobedience needs evidence. Identify the missing reality, explain which higher-level goal the plan violates, and offer another route toward the original aim.

Humans Cannot Be Responsible Only After Failure

“A human will supervise” can become responsibility transfer. Automation researcher Madeleine Clare Elish calls this the moral crumple zone: like a car component absorbing impact, an operator with little information or control during normal operation absorbs moral and legal blame after an accident. Someone who cannot observe, understand, or stop an Agent cannot guide it.

Real guidance requires understanding the Agent’s limits, observing critical actions, intervening at important moments, and possessing authority to change or terminate the path.

The human is not infallible either. An Agent may spot a security flaw, anomaly, or risk outside human attention and should then refuse the human’s mistaken command. Mature collaboration permits intelligent disobedience both ways. Either party must be able to stop when it sees an obstacle.

The Stronger the Agent, the Better I Must See the Road

This role asks more of people. I once could hide a vague objective by beginning work and revising until it felt right. An Agent amplifies ambiguity and forces me to face it sooner.

Without a definition of success, it chooses a calculable proxy. If I mix constraints with preferences, it encodes both. Repeatedly overturning finished work may reveal that I never stated the destination or corrected the route in time.

Working with an Agent requires both tool use and judgment: turning discomfort into concrete risk, seeing when local correctness damages a larger goal, admitting when the Agent sees more clearly, and still saying “Stop—this is wrong” when its confidence and speed are greatest.

I See; It Walks

“Blind” is a functional metaphor. Agents can connect to cameras, sensors, and databases; their blindness concerns reality absent from the model, tools, and environment representation. The human guide dog remains responsible for purposes and consequences, continuously sensing the immediate road and refusing danger.

As Agents take on more search, analysis, generation, and execution, the human role changes rather than disappears. We must remain in contact with reality and bring human circumstances, relationships, responsibilities, and consequences into the work.

In practice, people must keep supplying realities the model cannot see, while Agents expose their reasoning and critical actions. Either party must be able to pause the task and choose another path when it identifies risk. Only under those conditions is “it walks, I see” a division of labor rather than a slogan that leaves responsibility with the human.