Sakana Fugu Ships Single-Model Multi-Agent Orchestration via API

Sakana released Fugu with a one-line install for Codex on Ubuntu and macOS. The system presents Chat Completions and Responses endpoints while internally running TRINITY, an evolutionary coordinator that assigns three distinct roles to separate LLMs without weight merging. A second Conductor model trained via reinforcement learning generates communication topologies and instructions. Both components originate from ICLR 2026 papers and continue to receive retraining as new base models enter the pool.

Benchmarks in the June 2026 technical report (arXiv:2606.21228) show Fugu outperforming three anonymized frontier baselines on multi-step tasks. Gains derive from runtime topology discovery rather than larger single-model scale. This approach directly contradicts prevailing practice of maintaining separate planner, coder, and critic instances connected by external orchestration layers.

The architecture compresses agent coordination into weights that can be served as one inference target. Operational impact includes reduced latency from avoided inter-model handoffs and simpler deployment for users who previously managed LangGraph or AutoGen graphs. Continuous pool updates imply that performance tracks the underlying frontier rather than freezing at release.

Next milestones include public release of the Conductor training harness and expanded endpoint support for tool-calling schemas already present in the Responses API.