Kubernetes controller for managing AI agent workloads. Defines CRDs (Agent, AgentRun, AgentPlaybook, SkillCard, SkillCollection, LLMProvider) and controllers for composing and executing agent...
概览
Kubernetes controller for managing AI agent workloads. Defines CRDs under the konveyor.io API group and controllers for composing and executing agent workloads via Agent Sandbox. The controller follows the Tekton Task/TaskRun pattern: an declares what is available (skills, gateways (provider/model combinations), container image, prompt, typed parameters) and an supplies concrete values (gateway selection, parameter values, instructions) to trigger execution. The controller is domain-agnostic. It does not call Hub, Backstage, or any inventory system. Parameter values are opaque — the controller validates and passes them through. The creator of the AgentRun (UI, CLI, CI pipeline) resolves application metadata before creating the CR. - is a hard dependency for workload execution - stay in the entry point — the agent does not receive push credentials - are Agent Skills directories, delivered as ordinary OCI images mounted via ImageVolumes (K8s 1.
README
以下为原文 README 节选，完整内容请查看原文。
Agentic Controller
Kubernetes controller for managing AI agent workloads. Defines CRDs
under the konveyor.io API group and controllers for composing and
executing agent workloads via Agent Sandbox.
Overview
The controller follows the Tekton Task/TaskRun pattern: an Agent declares what is available (skills, gateways (provider/model combinations), container image, prompt, typed parameters) and an AgentRun supplies concrete values (gateway selection, parameter values, instructions) to trigger execution.
The controller is domain-agnostic. It does not call Hub, Backstage, or any inventory system. Parameter values are opaque — the controller validates and passes them through. The creator of the AgentRun (UI, CLI, CI pipeline) resolves application metadata before creating the CR.
CRDs
| CRD | Purpose |
|---|---|
| SkillCard | One skill or rule, from an OCI image, a git repository, or inline content. Assembled into /opt/skills/{name}/ at pod init, where {name} is the skill’s own frontmatter name. |
| SkillCollection | Group of skills. Points at an OCI image holding several, in which case the controller writes a SkillCard per skill it finds, or references existing SkillCards. |
| Gateway | LLM service endpoint serving one provider/model combination. |
| Agent | Template declaring available skills, gateways, container image, prompt, and typed parameters. |
| AgentRun | Execute a single Agent with specific values. Creates an Agent Sandbox. |
| AgentWorkflow | Ordered sequence of stages, each referencing an Agent. |
| AgentWorkflowRun | Execute a workflow. Creates AgentRuns sequentially per stage. |
Key design decisions
- Agent Sandbox is a hard dependency for workload execution
- Git credentials stay in the entry point — the agent does not receive push credentials
- Skills are Agent Skills directories, delivered as ordinary OCI images mounted via ImageVolumes (K8s 1.33+), git clones, or inline content
- Workspaces are ephemeral — git is the persistence layer
- ACP over HTTP (via
goose serve) provides real-time observability and human-in-the-loop interaction - Hub provides curated REST endpoints for the UI
See docs/adr/ for the full set of architecture decision records.
Authoring images and skills
- Create an agentic base image — extend
agent-base, build language images, publish them, and understand the harness contract. - Create a SkillCard image — author, validate, package, publish, and reference skills and rules.
Getting started
See docs/getting-started.md for a step-by-step guide to deploying the controller, configuring a Gateway with LLM credentials, and creating your first AgentRun.
End-to-end testing
See test/e2e/README.md for Kind setup and commands for
推荐工具
换一个关键词，或者移除筛选条件。
安装
npx skillfish add konveyor/agentic-controller