Price terms as published
Each line is the vendor’s own wording, split from the price capture checked on 2026-10-03. Nothing here is rounded, converted or re-worded — the full sentences stay in the review below. Lines that carry only a price continue the plan named on the line above, in the vendor’s own order.
| # | Plan / term | Checked |
|---|---|---|
| 1 | Free (Apache-2.0 open source | 2026-10-03 |
| 2 | no paid tier, no hosted plan on reefinfra.ai) | 2026-10-03 |
One-liner: Reef is Apache-2.0 Python infrastructure for agents that keep improving from their own use. It serves inference, records every interaction, matches feedback back to records, trains an update, evaluates the candidate, and publishes accepted updates as versioned releases - either model weights or the agent’s own harness (prompts, rules, skills).
Verified facts (checked 2026-10-03)
| Attribute | Value | Source |
|---|---|---|
| GitHub stars | 7,481 | Human-Agent-Society/reef via GitHub API, 2026-10-03 |
| Forks | 692 | GitHub API |
| Open issues / open PRs | 233 / 471 | GitHub search API, same day |
| License | Apache-2.0 | GitHub API license field + README badge |
| Created | 2026-08-31 | GitHub API |
| Last push | 2026-10-02 | GitHub API |
| Primary language | Python | GitHub API |
| PyPI package | reef-infra 0.1.2, requires Python 3.12+ |
PyPI JSON API, 2026-10-03 |
| PyPI downloads | 514 in the last month, 168 last week | pypistats.org API, 2026-10-03 |
| System dependency | git-lfs required (artifact/checkpoint functionality) |
README install note |
| Docs version | v0.1.2 | reefinfra.ai docs chrome, 2026-10-03 |
| Price | Free; reefinfra.ai has no pricing page (/pricing returns 404) |
2026-10-03 |
| Maintainers listed | 27 named contributors in the README | README “The Team” section |
The idea, stated by the authors
Reef’s README calls itself “the first open-source infrastructure for continual self-improving agents” — that is the project’s own claim, not an independent finding, and no third-party survey found here backs or disputes the priority claim. What is verifiable is the framing: Reef deliberately treats the problem as continual self-improvement rather than recursive self-improvement (RSI). The authors’ launch post is explicit: they chose the broader framing because RSI’s closed loop - AI building the system that produces its next version - is not required for the problem they want solved.
The loop is four steps, and the README maps each to a module:
- Serve - handle agent requests, record the interaction (
service/,runtime/). - Observe - match later feedback back to the recorded interaction, decide what is
eligible to learn from (
storage/records.py,train/processors/). - Grow - produce an update from eligible records via the deployment’s recipe
(
recipe/,train/). - Commit - evaluate the candidate, apply the selection policy, publish accepted
updates with version history (
train/evaluation/,artifact/,surface/).
The receipt mechanism is the piece that makes this usable rather than a research toy.
Requests carry an x-reef-scenario header and the response carries
x-reef-agent-record-id; you later POST /reef/report with a numeric score, textual
or structured feedback, and the receipts it judges. Feedback arriving late, or
arriving from a human clicking a thumbs-up, still lands on the right interaction.
Two surfaces: weights or harness
This split is the practical decision, because one path needs GPUs and the other does not.
| You want | Path | What you need |
|---|---|---|
| A model that gets stronger | Weight training (Slime + SGLang) | A trainable model, a supported GPU stack, usable feedback |
| Your agent’s rules/skills to self-improve | Harness optimization | A model endpoint, representative tasks, an evaluator - no local GPUs |
The built-in harness recipe is Reefine. You point Reef at any OpenAI- or Anthropic-compatible upstream (Ollama, vLLM, anything with an API key), create a scenario, and ask for a change in plain language: “when I ask you to fix a bug, reproduce it with a failing test first.” Reef runs a candidate, evaluates it, and commits or rejects it.
The cookbook names the recipes by task type: scientific discovery (TTT-Discover, Guidance-TTT), continual learning on a task stream (SAO, SDFT, SDPO, Meta-Harness, GEPA), and learning from real usage with no scores (OpenClaw-RL, SkillClaw, Reefine). Measured results pages exist for TriMul, circle packing, AIME 2025, IMOAnswerBench, CEO-Bench, Terminal-Bench and WildClawBench.
The caveats
- This is infrastructure, not a product. There is no UI, no dashboard, no
onboarding wizard. The quickstart is
uv pip install reef-infra, start a server, send a request, report a score. - Weight training means real hardware. The docs route you to a GPU-requirements page for the training deployment. The harness path is the cheap entry point.
- Very young and very busy. v0.1.2 on PyPI, 233 open issues against 471 open PRs, created five weeks before this check. Churn is the norm, not the exception.
- Python 3.12+ and git-lfs are hard requirements, which rules out some minimal environments and a lot of serverless/container setups.
- The benchmark results are the authors’ own. Every measured number is a recipe’s result page in the same repo. No third-party reproduction found as of 2026-10-03.
- Feedback quality is the whole game. Reef can only learn from records someone bothers to score. The OpenClaw-RL recipe exists precisely because most usage has no score attached.
FAQ
Is Reef free? Yes - Apache-2.0, and reefinfra.ai has no pricing page at all (the
/pricing path 404s) as of 2026-10-03. It runs on your own hardware.
Do I need GPUs? For the harness-evolving path, no: Reefine needs a model endpoint (an API or a local Ollama) and an evaluator. For weight training, yes.
How is this different from just fine-tuning? The loop is closed and versioned: requests are recorded with receipts, feedback is matched back, candidates are evaluated against a selection policy, and only accepted updates are published - with a release chain your harness can inspect. A single fine-tune run has none of that.
Can I use it with my existing agent? Reef exposes OpenAI- and Anthropic-compatible
endpoints (/v1/chat/completions, /v1/messages), so most clients need only a base URL
change plus the scenario header.
Sources
- https://github.com/Human-Agent-Society/reef - README full text (fetched 2026-10-03)
- https://api.github.com/repos/Human-Agent-Society/reef - stars, forks, license, dates, language (2026-10-03)
- https://api.github.com/search/issues?q=repo:Human-Agent-Society/reef - open issues 233 / open PRs 471 (2026-10-03)
- https://pypi.org/pypi/reef-infra/json - version 0.1.2, Python 3.12+ requirement (2026-10-03)
- https://pypistats.org/api/packages/reef-infra/recent - 514 monthly / 168 weekly downloads (2026-10-03)
- https://reefinfra.ai/docs/getting-started/quickstart/ - quickstart, receipt/report loop, recipe catalog (extracted 2026-10-03)
- https://reefinfra.ai/docs/user-guide/evolve-your-model/ - weight-training GPU requirements (extracted 2026-10-03)
- https://github.com/Human-Agent-Society - org page, sibling projects (AutonomousResearch accepted at COLM 2026, Harbor evaluation infra) (2026-10-03)
- https://x.com/ao_qu18465/article/2094867930081337730 - author’s launch post, the “continual self-improvement over RSI” framing (2026-09-01)