Why do AI agent teams get used for one narrow task at a time, instead of running a whole project together?
Why does industry apply swarm coordination only to individual tasks?
This explores why companies building AI agents tend to use multi-agent 'swarms' for narrow, self-contained jobs instead of whole workflows. The corpus has no direct survey of industry practice, but it does explain why the narrow scoping is a sensible response to how agent teams fail.
This explores why companies building AI agents tend to use multi-agent 'swarms' for narrow, self-contained jobs instead of whole workflows. The collection has no direct survey of what industry actually deploys. What it does have is a clear account of why agent coordination breaks down as work gets broader, and that account explains the caution well.
The sharpest practitioner view comes from Cognition (the company behind the Devin coding agent). Walden Yan argues that parallel subagents are fragile for a reason that is easy to miss: every action an agent takes carries assumptions it never writes down. Two agents working side by side on parts of the same project will quietly make choices that clash, and neither can see the other's reasoning Why do parallel subagents produce conflicting design decisions?. His fix is to return to a single agent working step by step, so the context stays continuous. That explains the pattern. Splitting work is safe when the pieces don't depend on each other's hidden decisions, which is true of narrow tasks and rarely true of whole workflows. Benchmark research shows the same problem at scale: as networks of agents grow, they agree too late or adopt a plan without telling their neighbors. They also accept what other agents tell them without checking it, so one mistake spreads through the team Why do multi-agent systems fail to coordinate at scale?.
The less comfortable finding concerns cost. When researchers break down why multi-agent systems perform well, about 80% of the difference comes from how many tokens the system spends, not from clever coordination. These systems use around 15× more tokens than a single agent, and once tasks are past roughly 45% accuracy, adding more coordination actually makes results worse Are multi-agent systems actually intelligent coordination or just token spending? How does test-time scaling work at the agent level?. As single models get stronger, the advantage of using a team shrinks further When do multi-agent systems actually outperform single agents?. From a business point of view, a swarm is often just an expensive way to buy more compute. That's worth paying for a bounded task where parallel attempts clearly help, and much harder to justify across an entire process.
There's an older way to see this too. Malone and Crowston's organizational models, written long before LLMs, show that every coordination design gives something up. Centralized control is cheap to coordinate but breaks when the central piece fails. Decentralized markets survive failures but cost a lot to coordinate. Hierarchies sit in between and leave capacity unused How do coordination architectures trade off efficiency against resilience?. No design wins everywhere, so keeping swarms inside small, well-defined tasks keeps those trade-offs contained. Teams also bring risks a single agent doesn't have, such as attacks that trick agents into handing work to each other in an endless loop Can agents be tricked into delegating work in circles?.
The corpus also argues that the narrow use won't last forever. Some real work needs different kinds of expertise, parallel effort and independent checking, and no single agent can organize all of that however capable it is Do single agents always hit organizational limits?. In long-running scientific research, self-organizing agent teams that keep competing hypotheses alive and share their failures beat centralized planners on the same budget Can decentralized teams outperform central planners in long-running science?. Other work aims to make teams easier to manage: separating team roles, coordination rules and decision methods into parts that can be swapped Can multi-agent teams swap roles, coordination, and algorithms independently?, or automatically switching off agents that aren't contributing Can multi-agent teams automatically remove their weakest members?. The takeaway is that the narrow scoping isn't timidity. It reflects a real problem with shared context and cost. Swarms will probably move beyond individual tasks first in exploratory work like science, where agents disagreeing and pursuing different ideas helps more than it hurts.
Sources 11 notes
Walden Yan argues multi-agent parallelism is fundamentally fragile: subagents embed unwritten assumptions in every action, so agents working independently converge on conflicting choices. The fix is reverting to single-threaded linear agents where context remains continuous.
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.
Research shows token usage explains 80% of multi-agent performance variance, systems use 15× more tokens than single agents, and coordination yields negative returns above 45% accuracy. Performance gains come from token distribution, not coordination sophistication.
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.
Empirical analysis shows MAS performance gaps narrow with stronger models, with SAS outperforming in many cases. Three formal defect types—node-level bottlenecks, edge-level overwhelm, and path-level error propagation—explain when single agents win.
Show all 11 sources
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.
Research identifies a novel MAS-specific attack that weaponizes cross-agent delegation to form task cycles, distinct from applying existing attacks like prompt injection to teams. The mechanism requires multiple agents and has no single-agent counterpart.
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.
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.
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.
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.
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
- Towards a Science of Scaling Agent Systems
- Drop the Hierarchy and Roles: How Self-Organizing LLM Agents Outperform Designed Structures
- How we built our multi-agent research system
- Single-Agent LLMs Outperform Multi-Agent Systems on Multi-Hop Reasoning Under Equal Thinking Token Budgets
- AgentsNet: Coordination and Collaborative Reasoning in Multi-Agent LLMs
- Scaling Behavior of Single LLM-Driven Multi-Agent Systems
- Why Do Multi-agent LLM Systems Fail?