Claude Code

User-Level Agents

Autonomous specialists loaded from ~/.claude/agents/ — available across every project

ai-servicers.com

What Are Agents?

Agents are autonomous specialists invoked via the Task tool to perform complex, multi-step operations independently. Unlike skills (which provide knowledge), agents can use tools, coordinate with other agents, and produce comprehensive deliverables on their own.

Skills vs Agents

  • Skills — Markdown how-to guides, invoked by Claude as context
  • Agents — Autonomous LLM instances with their own tool access
  • Agents can run independently and return results

User-Level Scope

  • Loaded from ~/.claude/agents/
  • Available in every Claude Code session
  • Invoked via subagent_type="pm"
  • Currently: 1 agent (PM)
The PM Agent

Project Manager

Infrastructure and DevOps project orchestration specialist. Coordinates between Architect, Security, Developer, and QA agents to deliver complex multi-component projects.

Core Expertise

  • Task decomposition and delegation
  • Multi-agent workflow coordination
  • Progress tracking and reporting
  • Risk identification and mitigation
  • Resource allocation and timeline management
  • Stakeholder reporting and documentation

Technical Profile

  • Model: Claude Sonnet
  • Tools: Read, Write, Grep, Glob, Bash
  • Invocation: subagent_type="pm"
  • File: ~/.claude/agents/pm.md

Coordinates With

  • Architect — system design decisions
  • Security — network and auth configuration
  • Developer — implementation and deployment
  • QA — testing and quality validation

PM Agent Workflow

How the PM agent manages a complex multi-agent project from start to finish

1
Analyze the user request and break it into logical subtasks
2
Write task breakdown to ~/projects/data/claudeagents/pm/tasks/{project}/
3
Invoke Architect agent for system design decisions
4
Invoke Security agent for network and auth configuration
5
Invoke Developer agent for implementation and deployment
6
Monitor progress by reading agent output files
7
Synthesize final report to ~/projects/data/claudeagents/pm/reports/{project}/

When to Use the PM Agent

Use PM When...

  • Task requires multiple specialist agents
  • Complex project with distinct design, security, and build phases
  • Need coordination across sequential tasks with dependencies
  • Want a single coherent report synthesizing multi-agent work
  • Project involves new service deployment end-to-end

Sequential vs Parallel

  • Sequential — Architecture first, then Security, then Developer (typical deployment)
  • Parallel — Research tasks or independent components run simultaneously
  • Iterative — Complex projects with feedback loops between agents

Example Request

"Deploy an analytics API with Keycloak authentication and PostgreSQL" — PM breaks this into architecture design, network/auth config, and implementation tasks, then delegates each to the right specialist.

User-Level Agents — Summary

1 User-Level Agent
PM Project Manager
4 Agents It Coordinates
Sonnet Model

The PM agent orchestrates, it doesn't implement. It delegates to Architect, Security, Developer, and QA agents — then synthesizes their work into a coherent project report the user can act on.

User-level agents live at ~/.claude/agents/ and are available across all projects, unlike project-level agents which are scoped to a specific infrastructure workspace.

ai-servicers.com