reef docs
Reef sits between an agent harness and its model provider. It records each request, connects later feedback to the response it judged, and can publish an updated model or harness without redeploying the harness. A recipe decides which feedback is useful and what Reef should update.
Read in this order:
- Get started: how to install reef, start the service, point a harness at it, and watch a version advance.
- Connect your harness: the wire contract,
covering requests,
x-reef-*headers, reports, training tensors, and reading an updated artifact back. - Define a recipe: what you write and what reef runs, the kinds of recipe, and a worked example whose code is executed by reef's test suite.
- Evolve your harness: the other path, which improves the agent's text instead of its weights and needs no GPU.
- Deploy and train: the service config, the training driver, weight sync and checkpoints, and day-two operations.
Reach for the reference pages when you need them. The recipe catalog lists what ships and when to pick each one, design explains why reef is shaped the way it is, and the glossary collects every term in one place.
If you are working on reef itself, read
development and
testing. Design records live in
rfcs/. The package-level READMEs under reef/ document the code for
the people who change it, rather than for people using reef.
historical/ holds the previous documentation set. It is being rewritten into
the pages above and is not part of the reading order.