Half-price runs, no pod
execution: batch sends the prompt to the vendor’s batch API (Anthropic, OpenAI or Gemini) at
about half the price and records the answer on the Agent. See batch runs.
AI agent subsystem
The cheapest, most boring way to run coding agents on Kubernetes. Create an Agent; the controller runs Claude Code, Codex or Gemini CLI and writes the result to its status.
What is the ai agent subsystem?
Section titled “What is the ai agent subsystem?”ai-agent-subsystem turns “run a coding agent” into a declarative Kubernetes operation. You
describe what the agent should do (a recipe), where it runs (a runtime template), and which
run you want (parameters). A controller reconciles each run into a Kubernetes Job, supervises it,
streams its output, and records the result on the resource’s status. Kubernetes stays the single
source of truth.
Why this one
Section titled “Why this one”A signed, pinned install
One install.yaml per release: images pinned by digest, cosign-signed, with SBOMs and SLSA
provenance. Non-root, read-only root filesystem, NetworkPolicy. No cert-manager.
Kubernetes is the API
No database, no backend service, no UI to host. Your RBAC, quotas, GitOps and audit logs already cover it.
You never touch D
The CRDs are the API. Typed TypeScript contracts are generated from the same source, so they cannot drift. D matters only to contributors changing the API.
The three resources
Section titled “The three resources”Everything is built from three Custom Resources that reference each other in a chain:
flowchart LR
AD["AgentDefinition<br/>the recipe"] -->|agentDefRef| ST["Station<br/>the runtime"]
ST -->|stationRef| AG["Agent<br/>one run"]
AG -->|controller creates| JB["Job → Pod"]
classDef ad fill:#2f5fd8,color:#fff,stroke:#16224f;
classDef st fill:#3f7be0,color:#fff,stroke:#16224f;
classDef ag fill:#5b93ec,color:#fff,stroke:#16224f;
class AD ad
class ST st
class AG ag
AgentDefinition
The recipe: prompt template, model, allowed tools, permissions, and output sinks.
Station
The runtime: a Pod template plus a recipe reference and run-history limits.
Agent
One run: a Station reference, parameters, and a status that tracks the lifecycle.
Controller
Watches Agents, builds Jobs, patches status, and prunes old runs.
Start here
Section titled “Start here” How the pieces relate The full relationship schema: data model and runtime composition diagrams.
Architecture The D monorepo: three binaries and a shared library, statically linked.
Install One kubectl apply of a signed release manifest stands up the whole subsystem.
Examples Worked recipes: a bug-fixer and a pure-text story writer.