TL;DR

A technical report on Fugu and Fugu-Ultra, language-model orchestrators trained to construct query-adaptive scaffolds for teams of heterogeneous LLM agents.

Why it matters

This is scaling through coordination rather than a single policy. The training system combines large-scale fine-tuning, evolutionary search, and reinforcement learning to decide how multiple agents should be assembled for each task.

Key findings

  1. 01

    Fugu generates task-specific agentic scaffolds instead of applying one fixed multi-agent workflow to every query.

  2. 02

    The report combines fine-tuning, evolutionary algorithms, and reinforcement learning with production infrastructure for multi-agent execution.

  3. 03

    Reported benchmark gains mix model quality, orchestration, tool use, latency, and inference budget, so they should not be interpreted as a single-model scaling law.

Scaling dimensions

Models, methods & benchmarks

Models
Fugu, Fugu-Ultra
Algorithms
Reinforcement learning, Evolutionary algorithms, Multi-agent orchestration
Benchmarks
SWE-Bench Pro, Terminal Bench, LiveCodeBench, GPQA-Diamond, Humanity's Last Exam, CharXiv Reasoning

Topics

Read the original sourcearXiv