| name | build-agent |
| description | Tutorial for building your first agent in the Agentics ecosystem. Covers Claude Code agents, MCP servers, and orchestrating multi-agent swarms. Use when a user asks how to build, design, or deploy an agent. |
Build Your First Agent
A walkthrough for building an agent that does useful work.
Choose the agent shape
Three common shapes:
| Shape | Use when | Example |
|---|
| Skill | Stateless guidance Claude follows | "How to deploy to GCP" |
| Slash command | One-shot user-triggered action | /events list |
| Subagent | Specialized model with own context | Code reviewer, research agent |
| MCP server | Programmatic tools (fetch, query, mutate) | agentics-org, agentics-events |
Most useful agents combine these: an MCP server provides tools, a skill teaches Claude when to call them, and a subagent owns a specialized workflow.
Minimal Claude Code agent (5 minutes)
plugins/my-agent/
โโโ .claude-plugin/plugin.json # name, version, description
โโโ agents/my-agent.md # frontmatter: name, description, model
โโโ skills/my-skill/SKILL.md # frontmatter: name, description
Then test with claude --plugin-dir ./plugins/my-agent.
Add tools with an MCP server
When you need real-world I/O (HTTP, files, APIs), wrap them in an MCP server:
- Add
.mcp.json to wire the server
- Write
server.js using @modelcontextprotocol/sdk
- Reference the tools in
allowed-tools: of your skill
The agentics-org plugin in this marketplace is a working reference. Read its server.js.
Orchestrate multiple agents
For complex jobs, spawn parallel agents using the Task tool:
- One message, multiple Task calls = parallel execution
- Each agent runs in isolation with its own context window
- Use a coordinator agent to merge results
See the swarm-orchestration skill for advanced patterns.
Output style
Lead with a working minimal example. Resist over-explaining theory. Drop them into a working scaffold first, then explain why each piece exists.
