Creating Custom Agents
CAMI includes agent-architect, a specialist agent that creates professional agents through conversation. You describe what you need, and it handles the rest.
When to Create a Custom Agent
Good reasons:
- You have a specific, repeatable process (deployment checklist, health checks)
- Your tech stack isn't covered by official agents (Django, Go, Flutter)
- Your team has unique patterns or conventions
Don't create one when:
- Official agents already handle your use case
- It's a one-off task (just ask Claude directly)
- You're not sure what you need (try official agents first)
Creating an Agent
Start Claude in your CAMI workspace and describe what you need:
cd ~/cami-workspaceclaudeYou: "I need an agent that checks Kubernetes pod healthfollowing our team's diagnostic workflow"Claude(via agent-architect): "I'll create a Workflow Specialistfor Kubernetes health checks. What are the specific steps inyour diagnostic process?"You: "First check if pods are running, then check resource usage,then verify endpoints are responding, then check logs for errors"Claude: "Got it. Let me structure this workflow..."[agent-architect creates the agent]Claude: "I've created k8s-pod-health-checker.md.Shall I save it to ~/cami-workspace/sources/my-agents/?"You: "Yes"Claude: "Saved! To use it, deploy to any project:'Deploy k8s-pod-health-checker to ~/projects/my-app'"What agent-architect Does
Based on your description, agent-architect:
- Classifies the agent - Determines if it's a Workflow Specialist, Technology Implementer, or Strategic Planner
- Gathers requirements - Asks targeted questions based on the class
- Generates the agent file - Creates proper frontmatter and methodology
- Saves to my-agents - Your personal agent collection
Agent Classes
agent-architect will classify your agent into one of three types:
| Class | Purpose | Examples |
|---|---|---|
| Workflow Specialist | Execute specific, repeatable processes | Health checks, deployment workflows, component builders |
| Technology Implementer | Build complete features in a domain | Frontend, backend, database, auth systems |
| Strategic Planner | Research, plan, and architect | Architect, security reviewer, researcher |
Tips for Good Agents
Be specific about what triggers it:
Good: "An agent for checking Kubernetes pod health in production"Vague: "A Kubernetes agent"Describe the workflow or domain clearly:
Good: "First check pod status, then resource usage, then logs"Vague: "Check pods and stuff"Mention tools or technologies:
Good: "Uses kubectl and our monitoring API at /health"Vague: "Uses standard tools"Using Your Custom Agent
Deploy it like any other agent:
You: "Deploy k8s-pod-health-checker to ~/projects/devops-tools"Claude: "Deployed k8s-pod-health-checker(v1.0.0 from my-agents)"# Later, in that project:You: "Check our production pod health"Claude(using your agent): "Running the Kubernetes health check workflow..."[Executes your defined steps]Iterating on Your Agent
Found something to improve? Just tell Claude:
You: "The k8s-pod-health-checker should also check database connections"Claude: "I'll update the agent to include database connection checks.Where should this fit in the workflow?"You: "After endpoint checks, before logs"Claude: "Updated and saved as v1.1.0"Sharing Custom Agents
Personal use: Agents in my-agents/ are already available
Team sharing: Create a Git repo and add it as a source:
You: "Add our team agents from git@github.com:ourteam/agents.git"Claude: "Added team-agents source with priority 50"Team members add the same source to get your shared agents.
Next Steps
- Agent frontmatter spec - Full specification for manual editing
- Agent architect methodology - Deep dive on agent design
- Team collaboration - Sharing agents across teams