Agent Team Builder

A meta-agent system for Claude Code that creates specialized AI agents through deep domain research and multi-lens quality review. Unlike template-based agent generators, this system researches the target domain before designing agents -- producing agents with real expertise, not generic instructions.

Why this exists

Most agent builders generate agents from templates or brief prompts. The result is generic agents that lack domain depth. Agent Team Builder takes a different approach:

Researches before designing. The architect agent runs 10-15 web searches, reads 8-10 sources, and extracts domain-specific competencies before writing a single line of the agent definition.
Reconnaissance before asking. A quick domain scan happens before user questions, so the system asks informed, domain-specific questions instead of generic ones.
Reviews before shipping. A dedicated reviewer agent evaluates every proposal through 7 independent lenses, catching overlap, missing interfaces, unrealistic workflows, and problem-solution misfit.
Creates agents that ask before acting. Every agent produced includes a domain-specific clarification step -- they never start working blind.

How it works

Two agents collaborate in a structured pipeline:

User request
    |
    v
[agent-architect]
    |-- STEP 0: Reads existing agent ecosystem
    |-- STEP 1: Quick domain reconnaissance (2-3 searches)
    |-- STEP 2: Asks informed, domain-specific questions
    |-- STEP 3: Deep research (10-15 sources per domain)
    |-- STEP 4: Extracts competencies from sources
    |-- STEP 5: Designs agent(s) with full spec
    |-- STEP 6: Sends proposal to reviewer
    |           |
    |           v
    |   [team-builder]
    |       |-- Reads ecosystem + researches real-world domain workflows
    |       |-- Evaluates through 7 lenses
    |       |-- Returns structured feedback
    |           |
    |           v
    |-- STEP 7: Incorporates feedback, creates files
    |-- STEP 8: Delivers summary
    v
Agent files in ~/.claude/agents/

Single and batch mode

Create one agent or an entire coordinated team in a single session:

Single mode: "Create a technical writing agent"
Batch mode: "Build me a team of 3 agents for my sales proposal workflow"

In batch mode, the system cross-checks between new agents, builds a responsibility table, maps inter-agent interfaces, and validates that the division into N agents (vs N-1 or N+1) is justified.

The 7 review lenses

Every proposal is evaluated through:

Lens	What it checks
1. Overlap	Competency conflicts with existing agents (and between new agents in batch)
2. Interfaces	Connection quality, handoff clarity, missing implicit interfaces
3. Skill granularity	Whether skills are truly atomic and reusable
4. Ecosystem balance	Domain bias, solo vs team ratio, user cognitive load
5. Problem-solution fit	Whether agents address the user's actual pain points and workflow
6. Workflow realism	Whether agent orchestration mirrors real-world domain workflows
7. Structural quality	Format compliance, tool references, checklist verifiability

What makes the output different

Every agent created by this system includes:

An identity statement (who the agent is, not what it does)
3-5 core philosophy principles
A strict-order process with tool-specific instructions
A domain-specific clarification step (asks before acting)
A verifiable quality checklist
Clear boundaries (what the agent does NOT do)
2-4 optional atomic skills

Installation

Copy both agents to your Claude Code agents directory:

git clone https://github.com/klausners/agent-team-builder.git
cp agent-team-builder/agents/*.md ~/.claude/agents/

Usage

In Claude Code, ask to create an agent or a team:

"Build me an agent for technical writing"
"I need a data engineering specialist agent"
"Create 3 agents to cover my content marketing workflow"
"Design a team that helps me build sales proposals"

Claude Code will automatically set up the architect + reviewer team and run the full pipeline.

Requirements

Claude Code with Opus model access
Internet access (for domain research via WebSearch/WebFetch)

Comparison with other approaches

Feature	Agent Team Builder	AutoAgent	Meta-Agent	Strands Agent Builder
Domain research before design	10-15 sources	No	No	No
Reconnaissance before questions	Yes	No	No	No
Independent quality review	7 lenses	No	Optimization loop	No
Batch/team creation	Yes, coordinated	Yes	Yes (FSM)	No
Real-world workflow validation	Yes	No	No	No
Created agents ask before acting	Yes, domain-specific	No	No	No
Platform	Claude Code	Python/API	Python/API	CLI

Structure

agents/
  agent-architect.md   # Researches domains, designs agents (single or batch)
  team-builder.md      # Reviews proposals through 7 lenses, validates workflow realism

License

MIT