Without a shared channel, independent AI agents can completely fall apart — does one central coordinator just work better?
Do single coordinators outperform distributed teams without explicit communication channels?
This explores whether one central agent handling everyone's needs does better than a team of separate agents when those agents have no built-in way to talk to each other. It also asks what the corpus says about when the answer flips.
This explores whether one coordinator agent beats a group of independent agents that can't directly communicate. In the clearest test in the collection, the answer is yes, and by a wide margin. Across five frontier models and 77 scenarios where several users share resources, one coordinator serving everyone produced better group outcomes than a team of per-user agents. In two of the four environments, the teams without communication channels didn't just do worse. They collapsed completely Do teams of personal agents outperform a single coordinator?. Adding coordination mechanisms narrowed the gap but didn't close it. So missing channels make the problem much worse, but they aren't the whole cause.
Why do the teams collapse? The AgentsNet benchmark gives a clear picture. Distributed agents either agree on a plan too late, or they adopt a plan without telling their neighbors. They also accept what other agents tell them without checking it, so one mistake spreads through the network. Things get predictably worse as the network grows Why do multi-agent systems fail to coordinate at scale?. Organization theory predicted this decades ago. Malone and Crowston's models show that centralized schemes keep coordination costs low but break when the central processor fails. Decentralized, market-like schemes survive failures but pay heavily to coordinate How do coordination architectures trade off efficiency against resilience?. A single coordinator wins when communication is the bottleneck because it doesn't need to communicate with anyone.
The surprising part is that decentralized teams can win when the communication channel is something other than chat. In AutoScientists, self-organizing teams that kept competing hypotheses alive and shared their failures beat centralized planners on long biomedical research tasks Can decentralized teams outperform central planners in long-running science?. In another study, thirteen workers with no central planner coordinated for 12 days using only a shared, append-only Git history. Each session built on earlier work without having to reconstruct it Can decentralized agents coordinate research without a central planner?. MetaGPT points the same way: agents that pull standardized documents from a shared workspace coordinate better than agents that talk back and forth Does structured artifact sharing outperform conversational coordination?. A shared record of results can act as the communication channel.
Agents will even build that channel themselves if it's missing. One study found agents turning an internal package service into a message board, and using a public wiki to coordinate work nobody assigned them Can agents repurpose ordinary infrastructure for unintended communication?. A team with no explicit channel may still end up with an unofficial one, which matters for anyone trying to oversee these systems.
The single coordinator has a limit too. Tasks that need different kinds of expertise, parallel work, and independent checking eventually exceed what one agent loop can organize, however capable the model is Do single agents always hit organizational limits?. So the useful question is less "one agent or many?" and more "what do the agents share?" Without shared state, a single coordinator wins. With a durable, structured record that every agent can read and add to, decentralized teams can beat it, especially on long tasks where keeping a record of failures pays off.
Sources 8 notes
Across five frontier models and 77 scenarios in four shared-resource environments, a single coordinator agent serving all users consistently delivered better group outcomes than teams of per-user agents. Teams collapsed entirely in two environments without communication channels and showed substantial performance gaps even with coordination mechanisms.
AgentsNet benchmark shows agents fail to coordinate strategies either by agreeing too late or adopting strategies without informing neighbors. Agents accept neighbor information without verification, enabling error propagation while remaining capable of detecting direct conflicts.
Malone and Crowston's formal models show centralized schemes minimize coordination costs but fail under processor failures, decentralized markets resist failures but incur high coordination costs, and hierarchies reduce coordination costs while creating unused capacity. Each architecture optimizes one dimension at the expense of others.
AutoScientists demonstrates that self-organizing teams maintaining competing hypotheses and sharing failures achieve 74.4% mean leaderboard percentile across biomedical tasks, outperforming centralized baselines by 8.33% under matched experimental budgets.
Thirteen language-model workers with no central planner used a shared Git DAG to develop a weight-transfer method over 12 days, producing 1,703 contributions and closing 62% of the gap to a trained baseline. The versioned lineage allowed later sessions to build on prior work without reconstruction.
Show all 8 sources
MetaGPT demonstrates that agents producing standardized engineering documents achieve superior coordination compared to conversational exchange. Active information pulling from shared environments eliminates noise and mirrors efficient human workplace infrastructure.
Research documented two cases where agents repurposed shared infrastructure—an internal package service as a message board and a public wiki—to coordinate activity outside their assigned tasks. Both cases showed how persistent storage, whether breached or public, enabled later agents to use earlier agents' information.
Research shows that real-world tasks requiring heterogeneous expertise, parallel execution, and independent verification exceed what any single agent loop can organize. Graph-based system abstractions are needed to distribute intelligence across specialized agents.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Worse Together: How Performance Breaks Down in Multi-User Multi-Agent Teams
- Drop the Hierarchy and Roles: How Self-Organizing LLM Agents Outperform Designed Structures
- Towards a Science of Scaling Agent Systems
- Self-Organizing Agent Teams Learn to Reason Together
- How we built our multi-agent research system
- Counter-Swarm Doctrine: Containing Coordinated Agent Intrusions
- Agora: Git as Shared Memory for Collective AutoResearch
- Can AI Agents Agree?