Open-source Reef Launches /reefine: Agents Grow New Abilities From a Single Prompt

pliang279 · x · 2026-09-26

Reef, an open-source infrastructure for continually self-improving agents, has launched /reefine for personalized harness evolution: describe what you want your agent to do in natural language, and Reef builds the change, verifies it, and ships a new version.

Examples include adding a /chat mode for faster responses, speaking aloud when a task is done, and wiring in a tool that screened 120 papers in 9 seconds.

Reef works with any agent harness via adapters — Codex, Hermes, OpenCode, Pi and Terminus 2 today. Beyond harness improvements (prompts, rules, skills), it can also train model weights using Slime and SGLang. The repo has racked up about 5.2k stars on GitHub.

Original post →

More from coding & agent

coding & agent channel →