Agent guidance reads like a CEO's messaging to their org

The frontier labs’ guidance on working with agents reads like a CEO’s messaging to their org.

Declare the values, state the goal, explain why it matters. Anthropic writes it plainly “generally favor cultivating good values and judgment over strict rules and decision procedures.” Netflix has said the same thing for fifteen years — context, not control.

Declare the values, state the goal, explain why it matters.

The Anthropic line comes from the constitution that shapes how Claude behaves. It also commits to explaining the reasons behind the rules it does keep. The Netflix idea goes back to the culture deck Reed Hastings shared in 2009, about seventeen years ago. The company’s culture page still asks managers for context, not control.

The everyday advice follows the same pattern. Anthropic’s prompting guide asks you to treat the model as “a brilliant but new employee who lacks context on your norms and workflows.” Its example is a bare rule, “NEVER use ellipses”. One sentence of reason works better: the response will be read aloud, and the text-to-speech engine won’t know how to pronounce them.

§Stating direction, reviewing output

Self-help books have been telling us to write down our objectives for decades and now we actually do, because that’s the only way to activate an AI agent / virtual employee. Everyone is quietly being promoted into a job that consists of stating direction and reviewing output and soon discover culture is paramount and try to address pathologies like:

1. Yes-men in people or Sycophancy in agents
2. Conformity in people or Prompt adherence in agents
3. Psychological unsafety in people or the “Make no mistakes” in agents
4. Silos in people or Agent sprawl in AI

Several of these pairs already have a documented machine version. In April 2025 OpenAI rolled back a GPT-4o update. It had become overly flattering and agreeable. Google’s Project Aristotle put psychological safety first among the factors behind its effective teams. The machine version is in the same prompting guide. On newer models, “CRITICAL: You MUST use this tool” causes overtriggering, and plain phrasing works better. The guide also warns that one model has “a strong predilection for subagents”. It spawns them where a direct approach would do.

We’re living in a new management age - where our instructions are executed by something with no ego, no career and no incentive to cover for you, and we can quickly find out how good they actually were.