reef vs sia
Continual-learning infra for self-improving agents: serve traffic (OpenAI/Anthropic-compatible), match feedback to runs, train weights via Slime/SGLang or evolve the harness, ship versioned updates. — versus — Self-improving loop from the SIA paper: Meta, Target and Feedback agents evolve a task agent's harness AND weights against a benchmark — #1 on MLE-Bench Hard, 14x kernel speedups.
Both implement a self-improving loop that evolves an agent's harness and weights; SIA is the paper's Meta/Target/Feedback agent recipe, Reef is production infra with serving, feedback matching and versioned delivery around it.
| reef | sia | |
|---|---|---|
| Stars | 911 | 2.1k |
| Forks | 57 | 256 |
| Language | Python | Python |
| License | Apache-2.0 | MIT |
| Last activity | yesterday | 16 days ago |
| Topics | training, agents | agents, training |
| Curated connections | 5 | 4 |
reef — the curator's take
Reef is the missing layer between an inference engine and an RL trainer: it serves live requests, ties feedback back to recorded interactions, runs a recipe (weight training via Slime, or prompt/rule/skill optimisation with no GPUs), evaluates candidates, and commits accepted versions with git-lfs history — all while staying online. Use it when you have a real feedback signal and want the agent to keep improving in production rather than in a one-off training run. Not for you if you just need a gateway or an eval harness, or if your model is a closed API and you have no harness surface to evolve. Weight-training path needs a supported GPU stack; harness-only path runs against any endpoint.
sia — the curator's take
The most ambitious of the self-improvement loops on the map: where others mutate prompts or code, SIA's Feedback agent rewrites the task agent's harness and updates its weights, with paper-grade receipts (70.1% on LawBench vs 45% prior SOTA, 14x on an AlphaFold-3 Triton kernel, #1 on MLE-Bench Hard). Reach for it when your problem IS a benchmark: a scoreable task the loop can grind against. NOT for tasks without a mechanical metric, and not a polished product — it's the official research implementation, so expect to build the task harness yourself and budget real GPU time for the weight-update path.