reef-infra 0.0.1

Continual learning infra for self-improving agents

Reef is infrastructure that serves an entire continual learning backend. Reef exposes standardized http endpoints so that you can download agents just like how you download codex or opencode using curl, and so that your agent can send its model requests to Reef's inference endpoint instead of the provider's.

The only difference is that, Reef constantly evaluates your agent behavior and improves the served harness and model weights in the backend. You keep getting better and better results without having to do anything.

Get started · Join Discord

Installation

From PyPI

pip install reef-infra

The distribution name is reef-infra; the Python package and CLI remain reef.

From source

git lfs install
git clone https://github.com/Human-Agent-Society/reef.git
cd reef
pip install -e .

Use the source checkout for development and for the training examples below, which also depend on the cookbook recipes and GPU runtime.

How it works

Reef processes each learning cycle in four steps. The table also shows which modules implement each step.

Using Reef

Reef supports two learning surfaces: model weights and agent harnesses. The recipe bound to a scenario determines which surface it updates.

1 · Serve

The following example starts the SAO (arXiv:2607.07508) example deployment. Run it from a Reef checkout in an environment that satisfies the GPU requirements in Evolve your model.

pip install -e ".[slime]" && pip install --no-deps --group runtime

export MODEL_PATH="Qwen/Qwen2.5-1.5B-Instruct"
export REEF_TOKEN="reef-local"

reef serve -c recipes/sao/examples/sao/serve.yaml \
  --reef.model_path "$MODEL_PATH" \
  --reef.port "8900"

curl -f http://127.0.0.1:8900/healthz          # ready to serve

2 · Train weights

Send inference requests through Reef and report a score for each response. The SAO recipe uses each eligible scored rollout to run a training step.

Send an inference request and report feedback

Reef's inference endpoint is OpenAI- and Anthropic-compatible: /v1/chat/completions and /v1/messages take the provider's own request body. The first request for a new scenario must include both Reef headers. Reef stores the recipe binding and rejects later attempts to change it.

The response body uses the provider's OpenAI-compatible format. Reef adds the x-reef-agent-record-id response header. Its value is the receipt that a later report uses to identify this interaction. A report can contain a numeric score, textual or structured feedback, and the receipts it evaluates. This example reports both a score and a short explanation.

import os
import httpx

reef = httpx.Client(
    base_url="http://127.0.0.1:8900",
    headers={"Authorization": f"Bearer {os.environ['REEF_TOKEN']}", "x-reef-scenario": "hello-reef"},
    timeout=300,
)

# Inference using Open-AI compatible format
response = reef.post(
    "/v1/chat/completions",
    json={
        "model": os.environ["MODEL_PATH"],
        "messages": [{"role": "user", "content": "Return exactly: reef is ready"}],
    },
)

receipt = response.headers["x-reef-agent-record-id"]
answer = response.json()["choices"][0]["message"]["content"]

# Sending report about the inference
matched = answer.strip() == "reef is ready"

reef.post(
    "/reef/report",
    json={"score": float(matched), "feedback": "matched" if matched else "wrong answer", "references": [receipt]},
).raise_for_status()

feedback carries the richer signal, plain text or a structured object, for recipes that read more than a scalar. The endpoint will validate the report schema (reef/core/reports/).

Watch it learn and grow

Once the recipe has enough feedback, it runs a training step and synchronizes the updated weights to the serving runtime. Later inference requests use the current version without restarting Reef.

3 · Evolve your harness

The harness_evolve recipe updates a harness tree that may contain rules, skills, configuration, prompts, and extensions. It builds a candidate from reported interactions, evaluates the current and candidate harnesses on the configured tasks, and publishes the candidate only when it wins that comparison. Harness scenarios do not share data or versions.

Install Reef harness that grows with you

You can install Reef harness like how you install most coding agents. The following is an example. A new scenario will be automatically created and bundled with the downloaded harness.

curl -fsS -H "Authorization: Bearer $REEF_TOKEN" \
  'http://localhost:8900/reef/harness/install?adapter=pi' | bash

reef-pi -p "fix the bug"

You can also retrieve an evolved harness by supplying its scenario in the header. For example, if you have a scenario harness-evolve-code-repair, you can install its harness via the following.

curl -fsS -H 'x-reef-scenario: harness-evolve-code-repair' \
  -H "Authorization: Bearer $REEF_TOKEN" \
  'http://localhost:8900/reef/harness/install?adapter=pi' | bash

Report a task result

reef-pi stores the receipts from a run, so its report command only needs the result you want to associate with the preceding interaction:

reef-pi -p "fix the failing test in auth.py"

# ... run your tests, grade the result ...

reef-pi report --score 0 --feedback "missed the empty-token case"
# reef-pi: reported 1 receipt(s) to harness-evolve-code-repair

Reef batches eligible reports according to the recipe configuration. When version checking is enabled, the adapter checks for a newer published version the next time it starts. Interactive sessions offer Update with … and Skip before accepting input; choosing update runs the installer directly. Headless sessions print the instruction instead. The harness evolution guide describes the proposal, evaluation, and publication process.

Cookbook recipes

Choose a recipe based on the feedback available from the workload and the artifact that should be updated. These implementations live in this repository's recipes/ cookbook; they are selected by dotted class reference and do not ship in the Reef wheel.

How is Reef different?

Reef builds the infra for AI that grows:

Learn more

The documentation is organized in the following order:

• Quickstart: install Reef, connect a client, and inspect the version history
• HTTP API: use the HTTP API and report feedback
• Write a recipe: configure how Reef processes data and produces updates
• Evolve your harness: evolve a harness instead of model weights
• Evolve your model: configure and operate a training deployment
• Recipes: additional references on the cookbook implementations in this repository
• Architecture: Overall architecture of Reef
• Glossary: Explanation of the terminologies used

Community & Contributing

Working on continual self-improving agent?

• Join Discord to share your recipes, ask implementation questions, and discuss new features.
• Start contributing with the contribution guide.
• Propose designs through an RFC issue.
• Report suspected vulnerabilities privately by following the security policy.

If Reef looks useful to you, please give it a ⭐ — it helps the community to discover and contribute to the project.