Meta's ReActNet compiles a fresh agent topology per query, cutting 20-agent runs from 7h to 6min
blaizedsouza · x · 2026-09-13
Meta engineers' ReActNet eliminates the most expensive part of multi-agent systems: instead of training a communication topology, it compiles a fresh one for every query, dropping 20-agent runtimes from 7 hours to 6 minutes.
The pain point: 5 agents works fine, but 20 turns into a group chat that's slower than one model and costs more than the task is worth.
How it works — no RL, no gradients, no training stage:
- An LLM controller reads the query and the agent roster
- It compiles a sequence of directed graphs, one per reasoning stage
- Every edge carries a written instruction (e.g. "list boundary cases for this behavior")
- Each agent updates its state from its own previous state plus assigned neighbors
- A final node aggregates all states into the answer
Scores 92.75 average on gpt-4o benchmarks.
More from coding & agent
- MiniMax Design plugs into Blender via MCP to drive AI video from 3D references — Hailuo_AI · 2026-09-14
- A TypeScript-to-Rust migration revives the classic Rewrite It In Rust meme — DanielLockyer · 2026-09-14
- gemini-cli PR fixes nested .gitignore trailing-slash patterns wrongly anchored — dylanyunlon · 2026-09-14
- Building a 'race your own ghosts' game with Gemini's Astra agent — LexicalLegend · 2026-09-14
- ChainForge's evalstats brings statistical rigor to LLM eval visualization after six months of work — IanArawjo · 2026-09-14
- RiftPoint: a Rust-backed multiversal checkpointer for LangGraph agents — goldenphoenix713 · 2026-09-14