Reef
Reef is open-source infrastructure that records agent inference and feedback, then uses them to update model weights or agent harnesses.
No visits from OpenBeat in the last 30 days yet. Every click on “Visit” is counted here.
Pricing and API checked by an OpenBeat editor on Nov 10, 2026 · facts from the maker’s own site ·
Tool insights
Reef sits between your agents and the model engine, logs every request, response and piece of feedback, and uses that record to improve the system. It can train new weights and swap them in live, or evolve the agent harness and keep the best version.
Agents usually stay as good as the day they were deployed, because their real usage never feeds back into the model or the harness. Reef gives teams one place to capture that usage and turn it into updates.
Teams running their own agents on open model stacks who want to learn from production traffic, and researchers building custom learning methods. It is a poor fit if you only call hosted models through an API and have no control over weights or harness.
What it does
Reef is open-source continual learning infrastructure from Human-Agent-Society. It serves live agent inference traffic through a versioned endpoint and records each request, response and feedback signal. It then uses that record to improve the agent in one of two ways: training model weights, or evolving the agent harness.
The main use case is an agent that already runs on real tasks and should get better from its own usage, without a manual retraining pipeline.
Key features
- Serves live agent inference through a versioned endpoint.
- Records every request, response and feedback item.
- Trains model weights from agent feedback.
- Hot-swaps new weights into the serving engine.
- Evolves the agent harness tree on your own tasks.
- Keeps the winning version as the active selection.
- Offers a Python API for custom learning methods.
- Offers an HTTP API for inference, feedback and artifacts.
Recent changes listed by the maker include reef serve, which starts inference without a YAML file. They also include Reefine, which refines a coding harness from plain-language instructions and, according to the maker, needs no GPU. Coral test-time training was added as a beta recipe.
Pricing
As of November 2026, Reef is open source under the Apache-2.0 license. The maker's site does not state a price, a paid plan or a hosted offering. Whether a free plan exists in the sense of a hosted tier is not stated. Costs of running your own serving and training infrastructure are not covered in the facts we have.
API and integrations
Reef has an HTTP API for inference, feedback and artifacts, and a Python API for writing your own learning methods. No separate SDKs are listed. The interface is a command-line tool (CLI).
Integrations named by the maker:
- SGLang
- slime
- Tinker
- pi
- opencode
- Claude Code
- Codex
- DeepSeek Harness
- Hermes Agent
- Terminus 2
The docs start at the getting-started page on reefinfra.ai, and the code is on GitHub.
Good to know
- Coral test-time training is a beta recipe, so expect it to change.
- The latest GitHub release is v0.1.3, dated 14 October 2026. The version number suggests an early project.
- As of November 2026, the GitHub repository shows about 7,800 stars, is not archived, and was last pushed on 9 November 2026.
- Founding date, customers and plan limits are not stated on the maker's site.
- Dates for most recent changes are not given in the facts, so check the repository for timing.
- Claims about learning gains from feedback are the maker's. We have no benchmark figures to check them.
FAQ
What is Reef?
Reef is open-source continual learning infrastructure from Human-Agent-Society. It records agent inference traffic and feedback, then uses that data to train model weights or evolve the agent harness.
Is Reef free and open source?
Yes, Reef is open source under the Apache-2.0 license as of November 2026. The maker's site does not state any paid plans or a hosted tier, and the cost of running your own infrastructure is not stated.
Does Reef have an API?
Yes. Reef provides an HTTP API for inference, feedback and artifacts, plus a Python API for custom learning methods. No dedicated SDKs are listed on the maker's site as of November 2026.
Which tools does Reef work with?
The maker names SGLang, slime, Tinker, pi, opencode, Claude Code, Codex, DeepSeek Harness, Hermes Agent and Terminus 2. Reef is used through a CLI.
Is Reef ready for production use?
The facts do not say. The latest release is v0.1.3 from October 2026, and the Coral test-time training recipe is explicitly marked as beta. Test it on a non-critical agent first.
Releases
No tracked releases yet.
In the news
No coverage yet.