Reef: open-source self-improving agent infra screens 120 arXiv papers in 9 seconds
alifcoder · x · 2026-09-26
- Reef (Human-Agent-Society, 5.6k GitHub stars) is the first open-source infrastructure for continual self-improving agents, connecting agent inference, feedback, learning, and versioned delivery.
- Demo: asking Reef to "add Jev as a tool" makes it evolve its own harness (prompts, rules, skills); the upgraded agent then hunted 2026 arXiv papers on self-evolving agents, screening 120 papers in 9 seconds.
- Supports training model weights via Slime and SGLang, or improving just the harness; agents improve by learning from user interactions.
More from coding & agent
- Open-source Agent Skill generates AGENTS.md files to guide AI coding agents — solyarisoftware · 2026-09-26
- Developer's detailed take: Opus 5.5 nears Fable 5.1 at lower cost, wins back Claude Code users — dejavucoder · 2026-09-26
- One prompt, nine minutes on a phone: motion graphics video built with Claude Code — henkvaness · 2026-09-26
- Source dive: DeepSeek Harness quietly relies on Pi libraries, with a 46-char system prompt — solyarisoftware · 2026-09-26
- exe.dev offers durable dev sandboxes with built-in coding agent Shelley — davidcrawshaw · 2026-09-26
- Claude 'wrote a bot' to beat all 50 game levels — survives only 4 minutes — banteg · 2026-09-26