If several AI agents work as a team, is 'who did it' still about one agent — or about how the team is organized?
Do multi-agent systems need different unit of analysis for agency?
This explores whether, once AI agents work in teams, we should stop treating the individual agent as the thing that 'acts' and start treating the team, its organization, or the infrastructure around it as the real source of what gets done and who answers for it.
This explores whether agency in multi-agent AI should be analyzed at the level of the team or organization rather than the single agent. The corpus gives a split answer: yes, the organization often becomes the right unit, but only when the interaction actually changes the outcome. A lot of what looks like 'multi-agent behavior' is just single-agent behavior repeated across several agents.
The case for shifting the unit comes from two sides. One side is about capability: real tasks that need different kinds of expertise, parallel work, and independent checking run into an organizational ceiling that no single agent loop can get past, however capable it is Do single agents always hit organizational limits?. The other side is about design. Older multi-agent research already treated the organization as its own object. MOISE+ describes a team through three separate layers: structure (roles and groups), function (shared plans), and rules about who is permitted or obliged to do what Can separating structure and function improve multiagent organization design?. Newer LLM frameworks such as IMACS revive this idea by making team roles, coordination style, and the method for merging answers into independent, swappable parts Can multi-agent teams swap roles, coordination, and algorithms independently?. If you can swap the coordination layer without touching the agents, then part of what the system 'does' lives in that layer, not inside any one agent.
The most interesting result is that agency at the organization level is real but hard to see unless you test for it. Labeling one agent as 'accountable' changes outcomes only when the workflow actually passes the deliverable through that agent Does accountability placement actually shape multi-agent outcomes?. On paper, responsibility does nothing. It only matters when it is built into how information flows. Something similar shows up in self-organizing teams. Setups with a fixed turn order but free choice of roles beat both strict top-down hierarchies and fully autonomous swarms. Agents in these setups invented specialized roles on their own and sat out tasks they weren't good at Do self-organizing agent teams outperform rigid hierarchies?. Who does what came out of the setup, not from any individual agent's design.
The counterweight is just as important. The security literature warns that putting several agents together doesn't by itself create a multi-agent effect. A failure only counts as multi-agent if the interaction amplifies it, creates it by combining agents, or produces a new property. A failure that stays the same is a single-agent problem in disguise Does a multi-agent setting automatically signal a security effect?. Performance work points the same way. The advantage of teams over single agents shrinks as models get stronger, and the failures that remain sit in specific places: a single agent becoming a bottleneck, the links between agents getting overloaded, or errors spreading along a chain of agents When do multi-agent systems actually outperform single agents?. Scaling studies find that the network layout can make error amplification 4 to 17 times worse, and that adding coordination stops helping once a single agent is already above roughly 45% accuracy When does adding more agents actually help systems?. AgentsNet shows a related weakness: agents accept what their neighbors tell them without checking it, so errors spread through the network even though each agent can spot a conflict when it sees one directly Why do multi-agent systems fail to coordinate at scale?. In each of these cases, the thing to analyze is the network layout or the connection between agents, not the agent.
The practical upshot reaches past engineering into governance. If agency is spread across a system, then attributing actions to specific agents, shaping how they interact, and fixing harm need infrastructure outside the agents themselves, much as internet protocols do for networked computers What infrastructure must exist beyond training agents?. Management scholars argue that monitoring has to come before any attempt to align incentives, because the system's decisions become harder to understand over time Must AI monitoring precede incentive alignment as agents evolve?. The answer the corpus supports is to use several units of analysis rather than replace one with another. Keep the agent as the unit for failures the team merely repeats, and treat the organization, its workflow, and its network layout as the unit for failures the team creates. The skill is telling the two apart.
Sources 11 notes
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.
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.
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.
Ablation studies show RACI accountability tags change outcomes only when the collaboration protocol's control flow passes the deliverable through the accountable agent. The best placement varies by model family, requiring per-model validation rather than universal design.
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.
Show all 11 sources
Interaction between agents can leave failures unchanged, amplify them, create them through composition, or define new properties. Only amplification, composition, and emergent properties qualify as genuinely multi-agent effects; unchanged failures reflect single-agent problems repackaged.
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.
Across 180 configurations, three dominant effects predict multi-agent success: tool-coordination trade-offs harm complex tasks, coordination stops helping above 45% accuracy, and topology choice controls error amplification by 4–17×. Architecture-task alignment, not agent count, determines outcomes.
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.
Chan et al. argue that external technical systems modeled on internet protocols are essential beyond behavior training. These systems must handle three core functions: attributing actions to agents, shaping their interactions, and detecting and remedying harm.
Humberd and Latham argue that as AI becomes more autonomous, monitoring mechanisms must be established first as prerequisites to incentive alignment, because defining desirable AI behaviors is harder than for humans and AI decision-making becomes increasingly incomprehensible over time.
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
- 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
- Towards a Science of AI Agent Reliability
- Embodied LLM Agents Learn to Cooperate in Organized Teams
- Agentifying Agentic AI