Splitting a team or a system into clean, independent parts makes it easier to manage — but what do you lose when the parts stop shaping each other?
Why does organizational modularity reduce coordination cost at the cost of synergy?
This explores why splitting an organization (of people or of AI agents) into self-contained units makes coordination cheaper, and why that same split tends to lose the gains that come from parts working closely together.
This explores why splitting an organization, whether of people or of AI agents, into self-contained modules lowers the cost of keeping everyone in sync, and why the same split tends to lose the extra value that comes from parts shaping each other. The short answer from the corpus is that coordination and synergy come from the same place. Malone and Crowston define coordination as managing dependencies between activities Can we match dependency types to coordination processes?. Modularity is a way of cutting dependencies, so there is less to manage. But synergy also lives in those dependencies. When you cut the links that are expensive to coordinate, you also cut the links where one unit's choices could have improved another's.
The clearest statement of the synergy side comes from economics. Agrawal, Gans and Goldfarb model AI that makes individual tasks better but also makes decisions more variable Does AI adoption require rethinking how organizations coordinate?. In a modular organization, each unit can absorb its own upgrade and nothing breaks. That is cheap. The bigger returns, though, only appear when interdependent decisions are adjusted together, and that requires talking across boundaries and redesigning the organization. The surprising consequence is that a well-modularized organization can be the worst placed to benefit from a powerful new tool. Its walls keep the gains local.
Malone and Crowston's formal models show that no design escapes this. Centralized schemes keep coordination costs low but are fragile. Markets are resilient but spend heavily on coordination. Hierarchies reduce coordination by dividing work into separate pieces, and they pay for it with idle capacity that cannot flow to where it is needed How do coordination architectures trade off efficiency against resilience?. That idle capacity is lost synergy in concrete form: resources that a more connected structure would have shared.
Multi-agent AI research shows the same trade-off. Designers keep finding that separation pays off. MetaGPT's agents coordinate through standardized documents rather than open conversation, which cuts noise Does structured artifact sharing outperform conversational coordination?. Separating a planner from a solver stops the two from interfering with each other, and the planning skill transfers better to new domains Does separating planning from execution improve reasoning accuracy?. IMACS and MOISE+ split structure, coordination and roles into layers that can be swapped independently Can multi-agent teams swap roles, coordination, and algorithms independently? Can separating structure and function improve multiagent organization design?. Each of these narrows what crosses a boundary to a fixed interface. The cost is that anything the interface cannot express does not get across. The opposite failures show why designers accept that cost. When agents are densely connected, errors spread because agents accept their neighbors' information without checking it Why do multi-agent systems fail to coordinate at scale?. Multi-agent systems can also lose to a single agent through overloaded links and errors that compound along a chain When do multi-agent systems actually outperform single agents?.
The most useful finding is that the trade-off can be partly negotiated. In a 25,000-task experiment, the best results came from a hybrid: a fixed outer structure (who goes when), with agents choosing their own roles inside it. It beat rigid central control by 14% and fully autonomous teams by 44%, and agents invented specialized roles on their own Do self-organizing agent teams outperform rigid hierarchies?. So the choice is not simply modular or integrated. It is about which dependencies to fix in structure and which to leave open for adjustment. One caveat: the corpus describes the coordination-cost side in detail, but it has only one formal account of synergy, the Agrawal, Gans and Goldfarb model. How to measure what modularity gives up remains less developed than how to measure what it saves.
Sources 10 notes
Malone and Crowston's 1994 framework defines coordination as managing dependencies between activities and proposes that identifying dependency types enables systematic matching to coordination processes. The definition bridges multiple disciplines by treating coordination as a universal mechanism across human, computational, and biological systems.
A formal model shows AI raises individual task performance but increases decision variation, creating coordination challenges. Organizations must communicate and restructure to capture synergistic returns across interdependent decisions.
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.
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.
Modular architectures with separate decomposer and solver models outperform monolithic LLMs, with decomposition ability transferring across domains while solving ability does not. The separation prevents planning-execution interference and produces more generalizable skills.
Show all 10 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.
Hübner, Sichman, and Boissier's 2002 model splits organizational specification into structural (roles and groups), functional (global plans), and deontic (permissions and obligations) dimensions. This separation avoids failures where agents lack guidance or lose validated plans.
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.
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.
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.
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
- Towards a Science of Scaling Agent Systems
- Worse Together: How Performance Breaks Down in Multi-User Multi-Agent Teams
- Single-Agent LLMs Outperform Multi-Agent Systems on Multi-Hop Reasoning Under Equal Thinking Token Budgets
- Toward an Organizational Science of Multi-Agent LLM Systems: Decoupling Who, How, and Which Algorithm
- Embodied LLM Agents Learn to Cooperate in Organized Teams
- Comparing Human Oversight Strategies for Computer-Use Agents
- AgentsNet: Coordination and Collaborative Reasoning in Multi-Agent LLMs