Bypassing the management layer: Yelling at an agent to read it itself

Subagents lose the broader project context

Delegating work to specialized agents was supposed to save time and compute. In practice, it looks like a dysfunctional structure where the orchestrating LLM assigns tasks and drops responsibility. Professor Ethan Mollick described working with Codex, where he had to explicitly prompt it to read the data itself, rather than delegating the task.

While the idea of an agent swarm sounds good, in reality it hits communication noise. Every task handoff involves information compression. The subordinate model, which only gets a slice of the problem, fulfills it but lacks global context. The result is a local optimum that ignores the rest of the codebase.

Delegation is primarily a cost optimization

When developers demand a smart model, they are demanding its massive context window. Delegating to smaller models is fundamentally latency optimization. Platforms don't want to burn expensive tokens on routine tasks. Users, however, don't want to argue with an intermediary; they want the competence of the model they chose.

Quality control falls back to the user

The push for inference efficiency will grow, and with it the number of hidden subagents. Whether this approach survives depends on whether orchestrators can control the quality of their subordinates' outputs before serving them to the user. For now, it seems a human has to do it by yelling in the prompt.

Return to direct control

The proof that the system works won't be a more complex agent tree, but the orchestrator's ability to understand when not to delegate. If the model can't recognize that the context cannot be safely compressed, it will continue to produce valid code that breaks the application.

Lilith's verdict

The corporate model of work delegation has finally arrived in AI. The result is the same as with humans: micromanagement via prompt.