Stateful inference infrastructure

Build agents that improve from every run

Reef records inference, connects feedback to the exact artifact that served it, and safely publishes the next version—without redeploying your harness.

Run Reef locally · 5 min See how Reef learns

The Reef loop

serve inference
  ↓
collect feedback
  ↓
train an artifact version
  ↓
publish safely
  ↺

Choose your path

Start with the outcome you need

New to Reef? Run the quickstart first. Already evaluating an integration? Jump directly to the contract or recipe guide.

Evaluate · 5 minutes

Run the first learning loop

Install Reef, send an inference request, attach feedback, and watch the artifact version advance.

Start the quickstart Integrate

Connect an agent harness

Learn the wire contract for inference, reports, version receipts, and reading an updated artifact back.

Open the integration guide Build

Define a learning recipe

Choose the signals, update behavior, and artifact type that Reef should produce for each scenario.

Design a recipe

How Reef learns

One traceable loop from request to artifact

Harnesses keep ownership of conversations, tools, and environments. Reef owns the versioned learning state behind inference.

Explore the architecture

1Resolve the scenario and its current artifact
2Serve a provider-native inference request
3Collect reports, events, and rewards
4Train, publish, and serve the next artifact version

Quickstart

From checkout to a healthy Reef service

Start locally, bind a scenario, and send the first provider-native request. The guide explains every step and expected result.

Follow the 5-minute guide

export REEF_TOKEN="$(openssl rand -hex 16)"
reef serve -c reef/configs/inference-local-sglang.yaml

curl http://localhost:8900/healthz

Go deeper

Keep the reference close when you need it

Understand