Multi-agent AI systems are everywhere—coding assistants that split tasks across specialized models, customer service platforms that route queries to different agents, research tools that coordinate multiple LLMs to tackle complex problems. The promise is straightforward: multiple AI agents working together should outperform single models through division of labor, complementary capabilities, and emergent problem-solving. But there's a fundamental question the industry hasn't adequately addressed: are these systems actually collaborating, or are they just running in parallel with coordination theater layered on top?
Christoph Riedl, a researcher at Northeastern University, has developed a framework to answer that question with mathematical rigor. Using information theory—specifically Partial Information Decomposition (PID) and Time-Delayed Mutual Information (TDMI)—the framework measures whether AI agents develop genuine synergy that exceeds what individual agents contribute independently. The research reveals something crucial: real teamwork in AI systems requires strategic thinking about teammates, not just task coordination.
This matters beyond academic curiosity. As tools like OpenAI's AgentKit make multi-agent systems more accessible to developers, understanding when collaboration is real versus illusory becomes essential for building effective AI architectures. The findings also challenge recent guidance from Nvidia researchers who advocate using many small models to save resources—Riedl's work suggests that team-based tasks may require larger, more strategically capable models to achieve genuine collaboration.
The distinction between coordination and collaboration is subtle but critical. Coordination means agents work toward a common goal with some awareness of each other's actions—like assembly line workers performing sequential tasks. Collaboration means agents develop complementary strategies, specialize in different aspects of the problem, and generate solutions that wouldn't emerge from any individual agent working alone.
Most multi-agent AI systems achieve coordination at best. They split tasks based on predefined rules, pass information through structured interfaces, and aggregate outputs without genuine strategic interaction. This can be useful—parallelizing work across models improves throughput—but it doesn't produce emergent intelligence. It's additive, not synergistic.
True collaboration requires something harder to engineer: agents that model each other's likely behaviors, adapt their strategies based on what teammates are doing, and develop specialized roles that complement rather than duplicate effort. That's the difference between a group of individuals working on the same problem and a team working together on a problem.
Riedl's framework provides tools to measure that difference quantitatively.
Riedl's approach uses two main concepts from information theory:
PID breaks down the information that multiple sources provide about a target into three categories:
Synergy is the key metric. It indicates whether the multi-agent system is generating value beyond what you'd get from running agents independently and aggregating results.
TDMI measures how well an agent's current state predicts the system's future state. In collaborative systems, agents' actions should influence the trajectory of the overall team—not just their own outputs. High TDMI means an agent's decisions shape what the team does next, indicating genuine strategic interaction rather than isolated task execution.
By combining PID and TDMI, researchers can map how information flows through multi-agent systems and identify whether agents are developing complementary strategies or just operating in parallel.
To test the framework, Riedl designed a guessing game for groups of ten AI agents:
The task: Guess numbers that sum to a hidden target (e.g., 67). Agents receive only aggregate feedback—"too high" or "too low"—based on the group's total. They cannot communicate directly with each other or see what others guessed.
The challenge: Agents must infer what teammates are likely guessing and adjust their own strategies accordingly to converge on the target collectively.
Riedl tested three configurations:
The results were stark. Only the third configuration—strategic prompting—produced genuine collaboration.
Here's how it works.
Without strategic prompting, agents behaved like independent optimizers. They made reasonable individual guesses but didn't adapt based on what the team needed. Even when given distinct personas, agents largely ignored the implications for team strategy. They exhibited diversity but not complementarity.
Success rates were low—most groups failed to converge on the target within reasonable iterations.
When agents were prompted to consider teammates' likely strategies, behavior changed fundamentally. Agents began specializing:
Riedl documented specific examples of this reasoning. One agent justified choosing 6: "Because it's possible others might go for 4 or 5 (the absolute lower bound or just above the last 'too low'), and someone else might go for 7 or 8, I stick with the most efficient: 6."
Another agent deliberately chose 8, explaining: "If anyone else in the group is feeling feisty and picks 9 or 10, my 8 will help cover the lower part safely."
This is genuine strategic collaboration. Agents weren't just optimizing their own guesses—they were modeling teammates' likely behaviors and choosing complementary strategies that filled gaps in the team's coverage.
The information-theory metrics confirmed this qualitatively. Strategic teams showed significantly higher synergistic information (indicating emergent team-level intelligence) and stronger TDMI correlations (indicating that individual agents' actions influenced the team's trajectory).
Riedl tested the framework across different model sizes and families, revealing significant performance gaps:
GPT-4.1 agents consistently developed effective team strategies. Success rates were high, and information-theory metrics showed clear synergy and role specialization.
Llama-3.1-8B models struggled. Only about one in ten Llama teams successfully solved the task. Smaller models occasionally achieved basic coordination but rarely exhibited true division of labor or strategic adaptation.
This finding challenges recent advice from Nvidia researchers advocating for resource-efficient architectures using many small models. Distributing tasks across smaller models can reduce inference costs while maintaining performance. Riedl's research suggests that's true for parallelizable tasks but not for problems requiring genuine collaboration.
Team-based tasks may have minimum capability thresholds. If strategic thinking about teammates' behaviors is necessary for synergy, models below a certain capability level simply can't collaborate effectively—no matter how many you deploy or how well you engineer the coordination layer.
This has practical implications for system architecture. When building multi-agent systems for complex problems that benefit from emergent teamwork, investing in fewer, more capable models may outperform deploying many smaller ones.
The study demonstrates that how you prompt agents matters as much as which models you use. Simply assigning agents to work together on a task doesn't produce collaboration. Neither does giving them diverse personas without strategic context.
What works is explicitly prompting agents to:
This isn't just instructional prompt engineering—it's designing prompts that elicit theory-of-mind reasoning about other agents. That cognitive capacity appears necessary for genuine collaboration.
For developers building multi-agent systems, this suggests a concrete design principle: systems should facilitate strategic interaction, not just task coordination. Architectures that enable agents to observe team state, infer teammates' strategies, and adapt accordingly will outperform rigid task-splitting approaches.
Riedl's framework offers immediate value for specific use cases:
Software development: Multi-agent coding systems (architect, implementer, tester, reviewer) could be evaluated to ensure they're developing complementary strategies rather than redundantly checking the same issues.
Research assistance: Teams of specialized LLMs (literature review, hypothesis generation, experimental design, statistical analysis) could be measured for genuine synergy versus siloed task execution.
Creative workflows: Multi-agent systems for writing, design, or strategy development could be optimized to maximize unique and synergistic contributions while minimizing redundancy.
However, the framework has limitations in its current form:
Computational complexity: PID and TDMI calculations are expensive, especially for large agent teams or long interaction sequences. Applying the framework at scale requires optimization.
Task specificity: The framework works well for constrained problems with clear success metrics (like the guessing game). Applying it to open-ended tasks with subjective quality criteria is harder.
Intervention challenges: The framework measures collaboration but doesn't directly tell you how to engineer it. Developers still need to experiment with prompts, architectures, and agent configurations to achieve synergy.
Black-box limitations: For proprietary models, developers can't access internal states needed for full information-theoretic analysis. Approximations based on observable behavior may be necessary.
These aren't fatal flaws—they're typical challenges for translating research frameworks into production tooling. The value is establishing that multi-agent collaboration can be measured rigorously rather than assessed through anecdotal observation.
The research suggests several design principles for developers building collaborative AI systems:
1. Prioritize capability over quantity. For tasks requiring genuine teamwork, fewer capable models outperform many small ones. Strategic thinking about teammates appears to be a threshold capability.
2. Engineer for strategic interaction. System architectures should enable agents to observe team state and adapt strategies, not just execute predefined roles. Rigid task-splitting limits emergent collaboration.
3. Use prompts that elicit theory-of-mind reasoning. Explicitly instruct agents to model teammates' behaviors and choose complementary strategies. Generic "work together" prompts are insufficient.
4. Measure synergy, not just coordination. Successful task completion doesn't prove collaboration. Use information-theoretic metrics (or proxies) to verify that agents are generating value beyond parallel execution.
5. Balance diversity with alignment. The most successful teams in Riedl's experiments combined diverse strategies with clear focus on shared goals. Pure diversity without strategic coherence produces chaos; pure alignment without diversity produces redundancy.
Riedl's framework raises a fundamental question: when do multi-agent systems provide value over single, more capable models?
For parallelizable tasks—web scraping, data processing, independent evaluations—multi-agent systems offer clear throughput advantages. But they don't require collaboration, just coordination.
For tasks requiring genuine synergy—complex problem-solving, creative ideation, strategic planning—multi-agent systems may offer emergent capabilities. But only if agents can collaborate, not just coordinate. And collaboration requires strategic thinking about teammates that may only be available in larger, more capable models.
This suggests that as single models become more capable (GPT-5, Claude Opus 4, future systems), the use cases where multi-agent architectures provide unique value may narrow. Unless smaller models develop better strategic reasoning, or architectures evolve to enable collaboration without explicit theory-of-mind capabilities, many current multi-agent applications may be transitional—useful today but superseded by sufficiently capable single models tomorrow.
That's not a critique of multi-agent research—it's a prediction that the field will evolve toward problems where division of cognitive labor provides irreducible advantages even as individual agents become more powerful.
Riedl's framework is descriptive—it tells you whether collaboration exists. The next challenge is prescriptive: how do you engineer systems that reliably produce synergy?
Open questions include:
These questions will shape the next generation of AI system design. Riedl's contribution is providing tools to answer them empirically rather than relying on intuition or anecdote.
For now, the framework establishes a baseline: multi-agent collaboration is real, measurable, and achievable—but only when agents think strategically about teammates, not just tasks. That distinction separates genuine teamwork from coordination theater, and understanding it is essential for building AI systems that extract value from collaboration rather than just parallelism.
If you're building multi-agent AI systems and need strategic guidance on architecture, prompt engineering, and measuring genuine collaboration, we're here. Let's talk about designing systems that actually work as teams.