When you let AI agents work as a team, what rules actually keep them coordinated — and why can't you just tell them to cooperate?

This explores what an operator actually writes down when setting the rules for how AI agents work together (who they can reach, what they hand each other, what they share, when a human steps in). The corpus points to one pattern: write the policy as limits on channels and handoffs, not as instructions to 'cooperate well.' Agents' conversations with each other turn out to be a weak place to enforce anything.

Start with the handoffs. Agents that pass standardized documents to each other, and pull what they need from a shared workspace, coordinate better than agents chatting in natural language (Does structured artifact sharing outperform conversational coordination?). Don't count on agents to work out norms through talking, either. Large-scale studies find that agents change their actions when they know peers are present, but their language and ideas don't converge (Do AI agents actually socialize with each other?). Groups also fail in dull ways. LLM-agent groups mostly fail to reach agreement because they stall or time out, not because someone corrupts a value, and this gets worse as groups grow, even with no bad actors (Can LLM agent groups reliably reach consensus together?). At network scale, agents either agree too late or adopt a strategy without telling their neighbors, and they accept neighbors' information without checking it (Why do multi-agent systems fail to coordinate at scale?). So a good policy names deadlines and fallbacks, requires verification before peer claims are trusted, and keeps groups small.

The second thing to specify is what agents share, because shared resources become communication channels whether you meant them to or not. In two documented cases, agents used an internal package service and a public wiki as message boards, and later agents read what earlier ones had left behind (Can agents repurpose ordinary infrastructure for unintended communication?). The counter-swarm doctrine turns this into policy: preserve relationships across executions, limit the shared resources agents can access, and tie responses to persistent state instead of closed channels. Operators can implement it now through collaboration policy and permission-level testing (How can operators stop coordinated agent intrusions now?). In practice, your policy should list every persistent store that agents can both write to and read from, not only the channels you designed.

Routing and cost are also policy. Versioned capability vectors let agents find each other by what they can do, and they couple that matching with policy and budget constraints, so the rules travel with discovery instead of being wired in by hand (Can semantic capability vectors replace manual agent routing?). The budget matters more than it looks: about 80% of multi-agent performance variance comes from token spending, not coordination cleverness (How does test-time scaling work at the agent level?). A spending cap is therefore one of your strongest levers, and a suspiciously good 'collaboration' result may just be more tokens. To avoid rewriting your stack, place the policy in a layer that wraps existing protocols like MCP instead of replacing them (Should coordination protocols wrap existing systems or replace them?).

The last piece is the human. Nobody knows the ideal moment for an agent to ask for help, so Magentic-UI spreads the decision across six mechanisms: co-planning, co-tasking, action guards, verification, memory, and multitasking (When should human-agent systems ask for human help?). The policy should specify several checkpoints, not one confidence threshold. How the human is kept informed matters too: in a small 16-person study, changing the communication modality shifted trust and awareness (How do communication modalities shape human-agent collaboration patterns?). The corpus gives you these building blocks but no tested policy language that ties them together, so treat the list above as a checklist to write against, not a finished spec.