Build your own agent.

Agentfile helps you build custom agents as portable container images. Declarative, harness-agnostic, and ready to run anywhere.

No code, declarative agents: Driven by Markdown and YAML and managed in git.
Bring your own harness: Leverage agentic harness tools you already know and trust - Claude, Codex, Pi, and more
Agents as portable container images: Standard container images that run anywhere - locally, in cloud, Kubernetes, or CI/CD.

bug-triage-agent/agentfile.yamlYAML


harness:
    claudecode: {}
llm:
    anthropic:
        model: claude-sonnet-5
systemPrompt:
    path: "prompts/systemPromptV3.md"
prompt:
    text: "your task is to investigate the bug ..."
skills:
    - git:
        url: "https://github.com/.../kubernetes-sre"
    - fs:
        path: "skills/known-issues"
mcps:
    - name: logs
      http:
        url: "https://internal-logs:1234/mcp"

What can you do with Agentfile?

Agentfile use cases

01Customize 02Automate 03Organize 04Operationalize

Don't use the same generic agent for everything.

AI agents are only as good as you make them. But how do you make an agent better? You give it skills, tools, and context.

CustomCreate custom agents specialized for each task.
ScopeEach agent carries its own skill set and configuration.
VarietyBuild an army of agents, and summon the right agent for the right task.

Build scriptable one-shot agents.

The prompt and goal are baked in. The harness is configured and tuned to run to completion. Standard input, standard output, and exit codes behave as you'd expect from any other process.

tail logfile.jsonl | af run log-triage
cd documents && af run grammar-checker --workspace .
cron "0 0 * * *" "af run daily-standup"

Organize skills around agents.

Move beyond one global pool of skills or duplicated repo setups. Each agent carries its own skill set and capabilities.

SharingShare and collaborate on your agent setup.
ControlAvoid cross-session, cross-skill contamination. You decide the context, not the agent.
ReuseNo duplication, no conflicts. Easily apply the right skills in the right projects.

Deploy agents on your terms.

Build Agents as standard container images. Deploy anywhere - your cloud, on-premises Kubernetes, GitHub Actions, or workflow engines.

af build -f myagent.yaml --tag my-agent:latest
docker push my-agent:latest
kubectl run my-agent --image=my-agent:latest

Ready to start building?

Check out the project on GitHub

Agentfile →