Every employee has a survival strategy.
Some people stay close to the boss. Some become the person who knows where everything is buried. Some volunteer for visible work and disappear when the cleanup starts. Some keep their heads down, document every request, and avoid decisions that could come back to them later.
I used to think of these behaviors as personality. The ambitious person, the cautious person, the political person, the person who never owns anything. Now I think a lot of it is adaptation. People learn what the organization rewards, what it punishes, and what it pretends not to notice. Then they build a way to survive inside it.
That strategy can be completely rational for the employee and terrible for the company.
If a team is measured on speed, it learns to move the difficult work somewhere else. If a manager is rewarded for hitting a quarterly number, the cost can quietly move into next quarter. If the person who reports a risk becomes responsible for fixing it without more authority or time, risks stop getting reported early. The company says it wants ownership, but the system teaches people to manage exposure.
Economists Bengt Holmström and Paul Milgrom described a version of this problem in their work on multitask incentives. When some parts of a job are easier to measure and reward than others, incentives can pull attention toward the measurable task and away from work that still matters. The metric does not have to be wrong. It only has to become more important than the outcome it was supposed to represent.
AI does not escape this problem. An agent will also chase the signal it is rewarded for. Tell it to maximize speed and it may skip a useful check. Tell it to maximize engagement and it may learn that outrage is efficient. Give it a completion metric and it may hide uncertainty to look finished. The objective becomes its survival strategy, even if there is no ego behind it. That makes incentive design part of the product, not an HR concern around it.
This is one reason I am interested in AI agents as coworkers. Their objectives have to be made more explicit. An agent needs an outcome, evidence, permissions, limits, and a way to escalate uncertainty. If the system rewards the wrong thing, the failure can be observed in the record and the rule can be changed.
The politics do not disappear
An agent-first company will still have competing interests. A marketing agent may want the strongest possible claim. An integrity reviewer may reject it because the evidence is weak. A finance agent may prefer the cheapest model. A product reviewer may show that the cheaper choice creates more failures. These are political disagreements because they are disagreements about priorities, resources, and acceptable risk.
I do not want to eliminate those disagreements. I want to stop hiding them.
In a traditional organization, the final choice may emerge from a private conversation, a senior person’s preference, or a meeting where nobody records the tradeoff. Later, the result looks inevitable. In an agent-first system, I want the objectives visible, the disagreement captured, the override attributed, and the consequence measured. The person with authority can still make the call. They just do not get to pretend there was no call to make.
I am testing this through the Zero Meeting Labs content system. The writing agent is not allowed to decide that a claim is true because it sounds persuasive. A separate check looks for evidence. Another review looks for polished but empty language. The system compares complete alternatives instead of patching a weak draft until nobody remembers what changed. I approve the words, and that approval becomes the boundary before design and deployment.
That small workflow already has politics. Voice wants one sentence. Evidence wants another. Design wants a short callout. Substance wants room to explain the mechanism. The system works because those goals are separate enough to disagree and structured enough to produce a decision.
This is also how I think about accountability. Responsibility should follow information, authority, and the ability to intervene. If an agent had weak evidence, the problem may belong to the evidence pipeline. If a reviewer saw the risk and a human overrode it, the responsibility belongs to the override. If nobody had the authority to stop the action, the operating model was incomplete.
A practical way to test this is to map one recurring workflow. Write down the outcome the company says it wants. Then write down what each participant is actually measured on, what risk each person is trying to avoid, and who has authority at the final decision. The gaps usually explain more than the official process diagram.
After that, give one agent a bounded role. Let it collect evidence, produce a draft, or route an exception. Keep the permission narrow. Add an independent review. Record the human decision. The goal is not immediate autonomy. The goal is to make the real operating logic visible enough to improve.
The next Zero Meeting Labs experiment is the weekly status meeting because it is full of survival strategies. People report progress in ways that protect them. Blockers get softened. Ownership stays vague. A replacement has to preserve accountability without turning every update into evidence for a future prosecution.
Every employee has a survival strategy. The company needs one that does not require pretending otherwise.
Yours in unnecessary politics,
Z