Buildkite execution environments

Make the environment sophisticated so the pipeline can stay simple.

Kait gives self-hosted Buildkite agents a known computational identity: the hardware, runtime, diagnostics, and workload capabilities are prepared before the job arrives.

The operating model

A pipeline expresses work. Buildkite routes that work. Kait supplies a reproducible capability. The workload executes where the declared hardware and runtime contract are true.

Start with the model

Buildkite as an execution substrate

A power-user perspective on pipelines, jobs, queues, and agents as the primitives of distributed execution.

Capability contract

How Kait turns runtime composition into validated workload capabilities such as data science, training, orchestration, and serving.

Architecture

The supervisor, container matrix, identity model, deployment paths, and release topology behind the contract.

Use it from Buildkite

Select the work capability and hardware in the agent selector. The pipeline requests intent; the worker advertises what its image can actually provide.

steps:
  - label: ":brain: train"
    command: "python train.py"
    agents:
      queue: ai
      kait.hardware: nvidia
      kait.capability.training: "true"

The selected container already contains the common execution substrate. The step does not spend its first minutes rebuilding Python, accelerator, and diagnostic machinery.

Direct destination or known-good base

Use an official Kait image directly, or derive an organizational image from an immutable Kait artifact and add only the internal delta. Platform teams maintain the common floor; application teams and agents bring the work that is unique to them.

View the repository See the pipeline examples