honeycomb

An AI agent fleet, as code. Seven Buzz agents build sediment, and sediment captures what they do.

How it works

• Agents are files. Config in agents/<role>.yaml, prompt in agents/prompts/<role>.md.
• One command deploys. scripts/apply.sh recreates only the pods that drift from the repo.
• Credentials stay in the cluster. Stored as K8s Secrets.
• Models swap at the gateway. LiteLLM routes each model_id to a provider.
• Sediment captures the work: completions, decisions, commit attribution, and edit survival.

Getting started

You need Buzz, a Kubernetes cluster, and a model gateway such as LiteLLM.

1. Build the agent image: CLUSTER_HOST=<cluster-host> scripts/build-image.sh, where <cluster-host> is the SSH name of the machine that runs your cluster.
2. Create each agent once in the Buzz app, on the Kubernetes provider.
3. Describe each agent in agents/<role>.yaml, starting from agents/capture.yaml.
4. Converge the fleet: scripts/apply.sh.

Stand up an agent covers each step, including the cluster names to change first.

Layout

agents/      one yaml per agent, with system prompts in agents/prompts/
image/       the agent image: pi harness, capture stack, git, gh, uv
manifests/   the pod template
relay/       the cluster host's relay config
scripts/     render, apply, build-image, sync-records, diff-relay
docs/        design, operations, and onboarding

Docs

• Agent fleet — design — why the fleet is shaped this way.
• Operate the fleet — stand up an agent, manage credentials and models, ship releases, and run the relay.
• Join the fleet — how members of our team connect to the fleet.

License

MIT