Should AI teams have a boss, or work it out among themselves — and does that answer change as the team grows?
Does hierarchical structure scale to larger embodied agent teams?
This explores whether putting a leader or chain of command in charge of a group of AI agents keeps working as the team grows, especially for agents acting in simulated physical spaces, or whether flatter, self-organizing setups do better at scale.
This explores whether a top-down chain of command keeps working as you add more AI agents to a team, particularly agents doing physical-style tasks such as household chores in a simulated home. One caveat first: the collection has only one study that deals directly with embodied teams. The answer below draws on that study plus broader multi-agent research, and none of it settles the embodied case at scale.
The embodied study points somewhere unexpected. In VirtualHome, a simulated house where agents split up chores, researchers didn't pick a team structure by hand. They had one LLM critique the team's performance and another rewrite its organizing instructions. The loop came up with chains, dual-leader setups, and teams that reorganized themselves mid-task, and these beat hand-designed hierarchies on both speed and communication cost Can LLMs automatically discover better team structures than humans?. So the useful question may not be "does hierarchy scale?" but "which shape of hierarchy fits this team at this size?" That answer seems to change as the team changes, and the system can search for it rather than fixing it in advance.
The larger non-embodied experiments point the same way: neither pure hierarchy nor pure anarchy wins. Across 25,000 tasks, several team sizes, and 8 models, the winner was a hybrid. An outside rule set the order of turns, but agents chose their own roles. It beat centralized control by 14% and fully autonomous teams by 44%, and agents invented specialist roles and sat out tasks they weren't good at Do self-organizing agent teams outperform rigid hierarchies?. A related approach scores each agent's contribution while the task runs and switches off the ones that aren't helping Can multi-agent teams automatically remove their weakest members?. For long-running scientific work, decentralized teams that keep competing hypotheses alive and share their failures beat central planners Can decentralized teams outperform central planners in long-running science?. The common thread is that a single boss tends to become a bottleneck and to narrow what the team explores.
The less obvious lever is what agents pass to each other, not who's in charge. MetaGPT found that agents trading standardized documents, like the specs and design docs a human engineering team uses, coordinate better than agents chatting Does structured artifact sharing outperform conversational coordination?. IMACS goes further and treats team roles, coordination method, and how results get merged as separate pieces you can swap independently Can multi-agent teams swap roles, coordination, and algorithms independently?. If that holds for embodied teams, scaling up might mean changing the coordination layer while leaving the roles alone.
There's also a skeptical reading. One analysis finds that about 80% of the variation in multi-agent performance comes from how many tokens the system spends, not from how cleverly it's organized How does test-time scaling work at the agent level?. Group deliberation also has its own failure modes, such as agents quietly converging on the same wrong answer What limits autonomous capability in large language models?. Before you credit a hierarchy (or its absence) for a bigger team's success, check whether the bigger team simply spent more compute.
Sources 8 notes
A Criticize-Reflect loop using LLM critique and reflection iteratively rewrites organizational prompts, discovering chain, dual-leader, and dynamic team structures that beat hand-designed alternatives on time and communication metrics in VirtualHome tasks.
A 25,000-task experiment across 8 models and multiple agent counts showed that sequential protocols with external ordering but internal role selection outperform centralized systems by 14% and fully autonomous systems by 44%. Agents spontaneously invented specialized roles and self-abstained when incompetent.
DyLAN's three-step importance scoring mechanism (propagation, aggregation, selection) quantifies individual agent contributions and automatically removes uninformative agents during inference, optimizing team composition without task-specific tuning.
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.
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.
Show all 8 sources
IMACS demonstrates that team structure (Belbin roles), coordination mechanism (Mintzberg framework), and fusion algorithm can be architecturally separated into independent components. This decoupling lets teams and protocols swap without rewriting each other, unlike existing frameworks that fuse these concerns together.
Research shows 80% of multi-agent performance variance comes from token budget, not coordination intelligence. LatentMAS and shared-KV-cache approaches offer ways to decouple performance gains from token costs.
Multi-agent deliberation produces specific failure modes (Degeneration-of-Thought, Silent Agreement), alignment at scale includes problematic self-valuation, and self-improvement is formally bounded by the generation-verification gap. Measurement error and conditional compliance hide the true capability ceiling.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Drop the Hierarchy and Roles: How Self-Organizing LLM Agents Outperform Designed Structures
- Worse Together: How Performance Breaks Down in Multi-User Multi-Agent Teams
- Towards a Science of Scaling Agent Systems
- Toward an Organizational Science of Multi-Agent LLM Systems: Decoupling Who, How, and Which Algorithm
- Embodied LLM Agents Learn to Cooperate in Organized Teams
- Self-Organizing Agent Teams Learn to Reason Together
- Scaling Behavior of Single LLM-Driven Multi-Agent Systems
- How we built our multi-agent research system