The coordination problem eating your AI budget

When one agent is not enough, the instinct is to add a supervisor: an orchestrator agent that tells the others what to do, checks their work and stitches the results together. It is the org chart, rebuilt in prompts, and it fails the same way org charts do — everything routes through the middle and the middle becomes the bottleneck.

The systems that work look less like hierarchies and more like teams with a shared whiteboard. Agents read from and write to a common state; they pick up work because it is there, not because they were told; and the 'supervisor' is reduced to a small set of rules about what is allowed.

This is also where most of the cost goes. Token spend in a chatty hierarchy is dominated by agents describing to each other what they just did. Give them memory and the narration disappears.

Design your multi-agent system around what the agents need to know, not who reports to whom, and both the reliability and the bill improve.