The Research Proves: MORE AI agents makes systems WORSE, not better
The Paradox of Adding More Agents
The Impact of Additional Agents on System Performance
Adding more agents to a system can lead to performance degradation, contradicting the common belief that increased computational resources improve outcomes.
Intuitively, if one agent completes a task in an hour, ten agents should finish it in six minutes. However, this assumption does not hold true in practice.
Increased agent count introduces coordination challenges: agents must wait for each other, duplicate efforts, and resolve conflicts. This overhead grows faster than the actual capability of the system.
A study by Google and MIT found that once a single agent's accuracy exceeds 45%, adding more agents results in diminishing or negative returns on efficiency.
Video description
My site: https://natebjones.com
Full Story w/ Prompt: https://natesnewsletter.substack.com/p/why-dumb-agents-mean-smart-orchestration?r=1z4sm5&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true
________________________
What's really happening with multi-agent AI systems? The common story is that more agents means more capability — but the reality is more complicated.
In this video, I share the inside scoop on why the frameworks get multi-agent architecture wrong:
• Why adding agents can actually make systems perform worse
• How serial dependencies block the conversion of compute into capability
• What Cursor and Yegge independently discovered about coordination collapse
• Why complexity should live in orchestration, not in agents
Google and MIT found that once single-agent accuracy exceeds 45%, adding more agents yields diminishing or negative returns. The team dynamics metaphor imports human coordination problems we've struggled with for centuries. The architectures that actually scale look almost too simple — two tiers, ignorant workers, no shared state, and planned endings.
For builders deploying agents at scale, the investment should go into orchestration systems — not into making individual agents smarter.
Chapters
00:00 The pitch for multi-agent systems is seductive but wrong
02:17 Core insight: simplicity scales, complexity creates serial dependencies
04:31 Google MIT study: more agents can mean worse outcomes
06:50 Rule 1: Two tiers, not teams
09:16 The team dynamics metaphor imports human coordination problems
11:34 Rule 2: Workers stay ignorant of the big picture
12:57 Rule 3: No shared state between workers
15:15 Rule 4: Plan for endings, not continuous operation
17:35 Yegge's Gastown universal propulsion principle
19:21 Rule 5: Prompts matter more than coordination infrastructure
21:42 Complexity lives in orchestration, not in agents
23:00 Why 10,000 dumb agents beats one brilliant agent
Subscribe for daily AI strategy and news.
For deeper playbooks and analysis: https://natesnewsletter.substack.com/
