Breaking: Sakana AI's RL Conductor Achieves State-of-the-Art Results with Minimal Compute
A tiny 7-billion-parameter language model trained via reinforcement learning is now outperforming AI giants like GPT-5 and Claude Sonnet 4 on complex reasoning and coding benchmarks. The model, called RL Conductor, automatically orchestrates a pool of larger worker LLMs – including GPT, Claude, and Gemini – without any human-written scripts.
According to Sakana AI researchers, the Conductor achieves these results at a fraction of the cost and with fewer API calls compared to manually designed multi-agent pipelines. The system is already powering Fugu, Sakana's commercial multi-agent orchestration service.
“While using frameworks with hard-coded pipelines like LangChain and Mixture-of-Agents can work well for specific use cases … In production, an inherent bottleneck arises when targeting domains with large user bases with very heterogeneous demands,” said Yujin Tang, co-author of the paper, in an interview with VentureBeat.
Tang emphasized that achieving real-world generalization in heterogeneous applications “inherently necessitates going beyond human-hardcoded designs.”
Background: The Limits of Manual Agentic Frameworks
Large language models possess powerful latent capabilities, but extracting their full potential has relied heavily on manually designed agentic workflows. Commercial AI products often use hard-coded pipelines like LangChain to chain together multiple models and steps.
These approaches are inherently rigid. As Tang explained, no single model is optimal for all tasks – one excels at scientific reasoning, another at code generation, math, or planning. Manually predicting and hard-coding the ideal combination for every query is practically impossible, especially with shifting query distributions.
“Every LangChain pipeline your team hardcodes starts breaking the moment the query distribution shifts — and it always shifts,” the researchers noted.
How RL Conductor Orchestrates an ‘Orchestra of Agents’
Instead of relying on fixed code or static routing, the RL Conductor generates a customized workflow for each input. It divides challenging problems, delegates targeted subtasks to the most suitable worker LLM, and designs communication topologies among agents.
The model is trained via reinforcement learning to dynamically analyze inputs, distribute labor, and coordinate responses. This automated coordination lets it adapt in real time to diverse queries without human intervention.
What This Means: Cheaper, More Flexible AI Pipelines
The RL Conductor's breakthrough has immediate practical implications. By replacing expensive, brittle hard-coded workflows with a lightweight learned system, companies can significantly reduce API costs while improving performance on varied tasks.
For developers, this means less time debugging brittle pipelines and more focus on high-level goals. The Conductor's ability to dynamically route tasks to the best model for each subtask also opens the door to more robust AI systems that handle real-world heterogeneity without constant manual tuning.
“An optimal agentic framework should be able to analyze a problem and delegate subtasks to the most suitable expert in the pool,” the paper states. RL Conductor appears to be the first practical implementation of that vision at scale.
Next Steps and Availability
Sakana AI has integrated the RL Conductor into its Fugu service, which is now commercially available. The research paper is also public, allowing the AI community to explore and build upon the approach.
The team plans to extend the Conductor to handle even more complex workflows and additional worker LLMs. If the trend continues, small learned orchestrators could become the standard way to harness the power of multiple large models.
Correction: An earlier version of this article misstated the benchmark results. The RL Conductor outperforms GPT-5 and Claude Sonnet 4 on the exact tasks tested, but broader comparisons are ongoing.