Agentihub Example

Example agent identities for the agenti* ecosystem.

This repo shows how to structure agent packages that agenticore can provision and run.

Three-Layer Architecture

agenticore   = execution engine (K8s pod — runs Claude Code, manages jobs)
agentihooks  = hook system + MCP tools (open-source — guardrails, integrations)
agentihub    = agent identities (CLAUDE.md, prompts, evaluation)

Agenticore provisions agents directly from this repo. On startup, agenticore's agent_mode/initializer.py clones the hub, finds the requested agent, and copies its package/ directory to /app/package/.

Directory Structure

agents/
└── <name>/
    ├── agent.yml              # Agent metadata (model, turns, timeout, MCP categories)
    ├── package/               # Production config — copied to /app/package/
    │   ├── CLAUDE.md          # Agent system prompt
    │   ├── system.md          # Extended system instructions
    │   ├── .claude/
    │   │   └── settings.json  # Claude Code settings (hooks, permissions)
    │   ├── .mcp.json          # MCP server config for the agent
    │   ├── prompts/           # Command prompts (optional)
    │   └── runners/           # Startup scripts (optional, sorted by prefix)
    └── evaluation/            # Eval harness
        ├── eval.yml
        └── cases/

Provisioning Flow

AGENTIHUB_URL + AGENTIHUB_AGENT env vars
          │
          ▼
    agenticore agent_mode/initializer.py
    git clone --depth 1 $AGENTIHUB_URL → /tmp/agentihub-clone/
          │
          ▼
    Copy agents/<name>/package/ → /app/package/
    Copy agents/<name>/evaluation/ → /app/evaluation/
          │
          ▼
    Claude Code runs with /app/package/ as config root

Example Agent: Publishing

The publishing agent demonstrates a content publishing pipeline that drafts, reviews, and publishes content to platforms like LinkedIn and Medium via MCP tools.

See agents/publishing/ for the full example.

Adding a New Agent

1. Create the agent directory:

   mkdir -p agents/myagent/package/.claude
   mkdir -p agents/myagent/evaluation/cases

2. Write agent.yml:

   name: myagent
   description: "What this agent does"
   
   claude:
     model: sonnet
     max_turns: 50
     permission_mode: bypassPermissions
     timeout: 1800
   
   mcp_categories: all

3. Write package/CLAUDE.md with the agent's system prompt and workflow instructions.

4. Create package/.claude/settings.json with permissions and hook wiring.

5. Set AGENTIHUB_URL and AGENTIHUB_AGENT=myagent in the agenticore pod environment.

Related Projects

Project	Description
agenticore	Claude Code runner and orchestrator
agentihooks	Hook system and MCP tools
agentibridge	Session persistence and remote control

License

MIT