Modern AI Engineering

Lesson 11.15 · 25 min · video lesson

Sakana Fugu: Lessons from an Open-Source Agent Study

What if, instead of training one giant model to be best at everything, we trained a model whose only skill is knowing which other model to ask, and how to make them work together?

In short: Sakana Fugu is an orchestration model from Sakana AI: a trained language model that sits in front of a pool of other LLMs and decides who should handle each request. Fugu routes each query to a single best worker using a small "selection head", trained first with supervised fine-tuning and then polished with evolutionary strategies. Fugu-Ultra writes whole multi-agent workflows in natural language, is trained with reinforcement learning (GRPO), and controls what each agent can see so agents do not simply copy each other.

What you will learn

  • What is Sakana Fugu?
  • Why Fugu was needed
  • The big picture: what Fugu does
  • Collective Intelligence
  • The two Fugus - Fugu and Fugu-Ultra
  • How Fugu picks the right model - the lightweight selection head
  • Teaching Fugu who is best - supervised fine-tuning
  • Polishing Fugu on real tasks - evolutionary strategies
  • How Fugu-Ultra conducts an orchestra - the Conductor
  • Teaching Fugu-Ultra to conduct - GRPO
  • Stopping the agents from copying each other
  • How well does Fugu perform?
  • The clever strategies Fugu discovered on its own
  • Quick Summary

In this lesson

  1. What is Sakana Fugu?
  2. Why Fugu was needed
  3. The big picture: what Fugu does
  4. Collective intelligence
  5. How Fugu picks the right model: the lightweight selection head
  6. Teaching Fugu who is best: SFT, then evolutionary strategies
  7. How Fugu-Ultra conducts an orchestra: the Conductor, GRPO and isolation
  8. How well does Fugu perform?
  9. The clever strategies Fugu discovered on its own
  10. Worked example, step by step
  11. Practice: try it yourself
  12. Quick summary

Key terms in this lesson

Orchestration model
A model trained to decide which other models to call and how to combine them.
Worker pool
The set of LLMs an orchestrator can delegate to.
Selection head
A small layer that maps the backbone's hidden state to one score per worker.
Soft target
A probability distribution used as a training label instead of a single correct class.
sep-CMA-ES
A scalable evolution strategy that searches parameters using only end-to-end scores, no gradients.
GRPO
Group Relative Policy Optimization: RL that scores each sample relative to the average of its group.
Intra-workflow isolation
Letting each agent see only its own work plus outputs explicitly granted by the access list.

This lesson is part of the Generative AI Engineering and Complete AI Engineer tracks. See plans and pricing to open the full lesson, its video, labs and quiz, or try the free lesson first (free account needed).

← 11.14 AI Orchestration: Coordinating Agents, Tools, and Flows · 11.16 Computer-Use Agents: Controlling Interfaces with AI →