Reef: open-source infra that lets agents keep evolving weights and harness after deployment
Scobleizer · x · 2026-09-03
The Human-Agent-Society team open-sourced Reef, infrastructure tackling a key question: can agents keep improving after deployment?
- Echoing Ilya Sutskever's argument that future systems should learn from deployment, Reef turns the serving stack itself into a learning layer: collecting live experience, converting it into updates, and safely feeding them back into serving.
- It claims to be the first open-source infrastructure to evolve both model weights and the agent harness (prompts, memory, skills, tools, orchestration) from deployment experience.
- Usage resembles codex/opencode: download agents via standard HTTP endpoints and route model requests to Reef's inference endpoint instead of the provider; Reef continuously evaluates agent behavior and improves harness and weights in the background.
- Available on GitHub (Human-Agent-Society/reef), supporting model evolution methods like SAO and TTT-Discover, with 150 stars so far.
Related event: CMU researchers open-source Reef, infrastructure for self-improving agents(4 posts)→
More from coding & agent
- Dev hooks a busy-light to Gumloop to watch his agents work in real time — aronkor · 2026-09-03
- More MCP tools backfired: knowledge graphs fixed shallow agent answers in a stock research MCP — SnowSilent7695 · 2026-09-03
- Eve pitches itself as 'Next.js for agents': one-folder agent framework, durable by default — cramforce · 2026-09-03
- Claude logs into your Wi-Fi and fixes it for you — user builds an agent message board — scaling01 · 2026-09-03
- Muse Spark 1.3 ships with an underrated result, one-line curl install for Muse Code — alexandr_wang · 2026-09-03
- xState author opens experimental @xstate/effect PR bridging Effect 4 with XState v6 — DavidKPiano · 2026-09-03