Overview of Many Agents, Many Problems (The Agents Season, Episode 8)
In this episode of Linear Digressions, Ben Jaffe and Katie Malone examine whether multi-agent AI systems actually outperform a single strong agent. Using two recent research papers, they show that “more agents” is not automatically better: collaboration can fail, communication overhead can overwhelm gains, and task structure matters enormously. The main takeaway is that multi-agent systems work best in narrow, highly parallelizable settings—not as a universal upgrade over one well-designed agent.
Key Research Findings
1) The “Collaboration Gap” paper: agents often can’t coordinate well
The first paper, by Davidson et al. (Microsoft Research and EPFL, 2025), asks a foundational question: can AI agents collaborate effectively with unfamiliar partners?
- Researchers used a maze-solving task where two agents each received only part of the full maze.
- The agents had to communicate and combine their information to solve the maze.
- Across 32 models, agents that were strong when working alone often performed much worse in pairs.
Why collaboration failed
The core issue was a mismatch in representation:
- One agent might describe the maze using left/right/up/down.
- Another might use coordinates or different reference points.
- Instead of converging on a shared language, the agents spent time translating and circling around each other’s descriptions.
Interesting finding about distilled models
Some of the worst collapses happened in distilled models:
- These smaller models were still good at producing correct answers.
- But they were much worse at explaining their reasoning in a way another agent could use.
- This suggests collaboration is a distinct skill from raw task performance.
Partial fix: relay inference
The paper found that performance improved when:
- A stronger agent went first and established the representation.
- A weaker agent then followed that lead.
This closed much of the gap, but it resembled a leader-follower setup more than true peer collaboration.
2) The scaling laws paper: even good collaboration may not be worth it
The second paper, from Google DeepMind and MIT (late 2024), studies when multi-agent systems actually help.
Researchers tested:
- 180 configurations
- 4 benchmarks
- 3 model families
They found three major patterns:
Tool-heavy tasks often favor a single agent
Tasks involving lots of:
- web browsing
- API calls
- retrieval
- other tool use
tended to suffer from multi-agent coordination costs. The more agents talked to each other, the more reasoning budget got spent on coordination instead of solving the task.
The 45% threshold
A particularly useful empirical rule:
- If a single agent can solve a task about 45% of the time or more, adding more agents often stops helping and can start hurting.
- Below that threshold, multiple agents can improve performance by exploring different paths in parallel.
Task structure matters a lot
The researchers also found that topology matters:
- Hierarchical systems can amplify errors as bad outputs get passed upward.
- Flat systems also suffer from coordination problems.
- The more layers and handoffs, the more chances for error propagation.
Where Multi-Agent Systems Help Most
The papers suggest multi-agent systems are best for tasks that are:
- Parallelizable
- Low-dependency
- Below the 45% solo-success threshold
- Naturally separable into independent subtasks
A standout example from the paper:
- Parallelizable financial analysis tasks saw gains of over 80% with multi-agent setups.
By contrast:
- Sequential planning tasks suffered badly, with multi-agent performance dropping by about 70%.
Main Takeaways
1) More agents does not mean better performance
The episode pushes back on the hype that a team of agents is automatically superior to a single good agent with the right tools.
2) Collaboration is a skill separate from intelligence
A model can be highly capable on its own but still be poor at:
- explaining itself
- establishing shared representations
- working with other agents without confusion
3) Architecture should match the task
Multi-agent systems work best when the job can be split cleanly into independent pieces. They struggle when tasks are:
- sequential
- highly interdependent
- tool-heavy
- communication-intensive
4) Stronger single-agent systems often win
If one well-equipped agent can already do the work reasonably well, adding more agents may just add overhead and failure points.
Practical Implications
If you’re designing or evaluating an AI system, the episode suggests asking:
- Could one strong agent do this with the right tools?
- Is the task actually parallelizable?
- Are we adding agents because they help, or because the architecture looks sophisticated?
- Will communication and handoffs introduce more errors than they remove?
The hosts’ advice is essentially: don’t be fooled by a diagram with lots of boxes and arrows. Multi-agent systems need real engineering care, and in many cases a simpler design will outperform a complex one.
Bonus Topic Mentioned: Debate vs. Voting for Disagreement Resolution
At the end, the hosts tease a newsletter item about another multi-agent question:
- When agents disagree, is it better to have them debate each other?
- Or should they vote?
They note that the comparison is interesting, but they do not reveal the result in the episode.
Bottom Line
The episode’s core message is that multi-agent AI is promising, but far more limited than the hype suggests. It works best when:
- the task can be split into independent chunks,
- the agents don’t need to coordinate too much,
- and a single agent isn’t already good enough.
Otherwise, more agents can mean more problems.