The $200 billion AI model race just hit a ceiling named "manual orchestration."
The Summary
- Sakana AI built RL Conductor, a 7B parameter model trained via reinforcement learning to automatically route tasks across frontier LLMs like GPT-5, Claude Sonnet 4, and Gemini 2.5 Pro
- The tiny orchestrator beats individual frontier models and hand-coded pipelines on reasoning and coding benchmarks while using fewer API calls and lower cost
- The system powers Fugu, Sakana's commercial orchestration service, addressing the core bottleneck: hardcoded pipelines break when query distributions shift
The Signal
We've been building the AI economy backwards. Companies spent billions training frontier models to be everything to everyone. Then they discovered the models couldn't actually deliver on heterogeneous real-world tasks. So they built LangChain pipelines and multi-agent frameworks, hardcoding decision trees that route queries to the right model at the right time.
Those pipelines worked fine in demos. In production, they shattered. Query distributions shift. User demands evolve. A pipeline optimized for last quarter's workload becomes this quarter's bottleneck. Sakana's solution is a 7 billion parameter model that learned to orchestrate itself.
"Achieving real-world generalization in heterogeneous applications inherently necessitates going beyond human-hardcoded designs."
RL Conductor treats the orchestration problem as a reinforcement learning task. The model analyzes incoming queries, dynamically distributes work across a pool of specialized LLMs, and coordinates their outputs. No human ever writes a routing rule. The conductor learned what GPT-5 is good at, where Claude excels, when to parallel process, when to chain reasoning.
The results matter because they're economic, not just technical:
- Beats individual frontier models on reasoning and coding benchmarks
- Outperforms expensive human-designed multi-agent systems
- Uses fewer API calls, cutting compute costs
- Adapts automatically when workload patterns change
This is the agent economy's infrastructure moment. Building agents isn't the hard part anymore. Orchestrating them at scale, across shifting distributions, without constant human intervention, that's the unlock. Sakana isn't selling a better model. They're selling the meta-layer that makes every model better by knowing when to use which one.
The timing aligns with a broader pattern. As frontier models commoditize, value migrates to the orchestration layer. OpenAI, Anthropic, and Google compete on raw capability. But in production, a smart 7B router that knows when to call which API might deliver more value than another 10x scale-up in parameters.
Co-author Yujin Tang identified the exact failure mode: large user bases with heterogeneous demands break hardcoded systems. That's not a bug in current frameworks. It's the architecture. LangChain and Mixture-of-Agents assume you can predict the distribution. You can't. Markets change. Users evolve. A system that can't adapt is a system that degrades.
The Implication
If you're building on LLM APIs today, your orchestration layer is now your moat. The models themselves are rented commodities. How you route between them, how you adapt to shifting workloads without rewriting code, that's defensible.
Watch for consolidation in the orchestration space. Sakana's Fugu is commercial now. Expect competitors within quarters. The companies that win Web4 won't be the ones training the biggest models. They'll be the ones that make every model smarter by knowing when to shut up and delegate.