Claude Agent Framework

A production-ready multi-agent orchestration framework built on Claude Agent SDK. Design, compose, and deploy complex AI workflows with pre-built architecture patterns.

中文文档

Overview

Claude Agent Framework addresses the fundamental challenge of complex tasks that require diverse specialized capabilities—research, analysis, code generation, decision-making—which cannot be effectively handled by a single LLM prompt. The framework decomposes these tasks into coordinated workflows where a lead agent orchestrates specialized subagents, each with focused prompts, constrained tool access, and appropriate model selection.

Built on Claude Agent SDK, it provides:

• 7 Battle-tested Patterns — Research, Pipeline, Critic-Actor, Specialist Pool, Debate, Reflexion, MapReduce
• Role-based Architecture — Separate role definitions from agent instances for flexible configuration
• Two-layer Prompt System — Framework prompts + business prompts for reusable workflows
• Production Plugin System — Lifecycle hooks for metrics, cost tracking, retry handling
• Full Observability — Structured JSONL logging, session tracking, debugging tools
• Simple API — Go from concept to working system in minutes

from claude_agent_framework import create_session

session = create_session("research")
async for msg in session.run("Analyze the competitive landscape of AI coding assistants"):
    print(msg)

---

Why Multi-Agent?

Complex tasks require multiple specialized capabilities. A research task needs web searching, data analysis, and report writing—each requiring different tools, prompts, and models. A single LLM prompt cannot effectively handle this.

The Problem with Monolithic Approaches:

• Prompt bloat: One prompt trying to do everything becomes unwieldy
• Tool overload: Agent accesses tools it shouldn't use at certain stages
• Quality degradation: Generalist prompts underperform specialized ones
• Cost inefficiency: Expensive models used for simple subtasks

The Solution: Agent Specialization and Orchestration

User Request
      │
      ▼
┌─────────────────────────────────────────────────────────────────┐
│                         Lead Agent                               │
│  • Receives request, decomposes into subtasks                    │
│  • Can ONLY use Task tool to dispatch sub-agents                 │
│  • Never performs research/analysis/writing itself               │
└────────────────────────────┬────────────────────────────────────┘
                             │ Task tool calls (parallel)
                             ▼
┌─────────────────────────────────────────────────────────────────┐
│  ┌─────────────┐  ┌─────────────┐  ┌─────────────┐              │
│  │  Worker 1   │  │  Worker 2   │  │  Processor  │  ...         │
│  │  (haiku)    │  │  (haiku)    │  │  (sonnet)   │              │
│  └──────┬──────┘  └──────┬──────┘  └──────┬──────┘              │
│         │                │                │                      │
│         └────────────────┼────────────────┘                      │
│                          ▼                                       │
│                   files/ (filesystem)                            │
│              Loose coupling via files                            │
└─────────────────────────────────────────────────────────────────┘

Key Principles:

Principle	Implementation
Separation of Concerns	Lead orchestrates, subagents execute
Tool Constraints	Each agent gets only the tools it needs
Loose Coupling	Filesystem-based data exchange
Cost Optimization	Match model capability to task complexity

---

Quick Start

pip install claude-agent-framework
export ANTHROPIC_API_KEY="your-api-key"

from claude_agent_framework import create_session
import asyncio

async def main():
    session = create_session("research")
    async for msg in session.run("Analyze the competitive landscape of AI coding assistants"):
        print(msg)

asyncio.run(main())

---

Architecture Patterns

The framework provides 7 pre-built architecture patterns:

Architecture	Pattern	When to Use
research	Master-worker parallel	Gathering data from multiple sources, market research, literature reviews
pipeline	Sequential stages	Code review, content workflows, multi-step approvals
critic_actor	Generate-evaluate loop	Quality improvement, iterative refinement
specialist_pool	Expert routing	Technical support, Q&A systems, diagnostics
debate	Pro-con deliberation	Decision support, risk assessment, vendor evaluation
reflexion	Execute-reflect-improve	Debugging, root cause analysis, optimization
mapreduce	Parallel map + reduce	Large-scale analysis (500+ files), batch processing

Each architecture defines roles with specific constraints:

# Research architecture roles
worker      → RoleCardinality.ONE_OR_MORE   # Parallel data gatherers
processor   → RoleCardinality.ZERO_OR_ONE   # Optional data processor
synthesizer → RoleCardinality.EXACTLY_ONE   # Required result aggregator

---

Building Business Applications

Building a business application involves 5 steps:

Step 1: Choose Architecture

Match your workflow pattern to an architecture:

• Parallel data gathering → Research
• Sequential stages → Pipeline
• Iterative improvement → Critic-Actor
• Expert routing → Specialist Pool
• Decision analysis → Debate
• Trial-and-error → Reflexion
• Large-scale processing → MapReduce

Step 2: Define Agent Instances

Map your business roles to architecture roles:

from claude_agent_framework.core.roles import AgentInstanceConfig

agents = [
    AgentInstanceConfig(
        name="market-researcher",      # Business name
        role="worker",                 # Architecture role
        description="Market data collection",
        prompt_file="researcher.txt",
    ),
    AgentInstanceConfig(
        name="report-writer",
        role="synthesizer",
    ),
]

Step 3: Write Business Prompts

Create prompts/ directory with business-specific context:

prompts/lead_agent.txt

# Competitive Intelligence Coordinator

You are coordinating analysis for ${company_name}.

## Team & Skills
- **Researchers**: Use `competitive-research` Skill for methodology
- **Report Writer**: Uses `report-generation` Skill for formatting

## Deliverables
- Research notes → files/research_notes/
- Final report → files/reports/

The framework merges your business prompts with generic framework prompts:

Final Prompt = Framework Prompt (generic role rules)
             + Business Prompt (domain context)
             + Template Variables (${company_name} → "Acme Corp")

Step 4: Create Skills (Optional)

Skills provide methodology guidance that agents invoke based on context:

.claude/skills/competitive-research/SKILL.md

---
name: competitive-research
description: Competitive intelligence methodology
---

# Research Focus
- Product & Service Analysis
- Market Position
- Financial Indicators

# Data Collection Priority
1. Market Data (size, share, growth)
2. Financial Data (revenue, valuation)
3. Technical Metrics

# Output Specification
Save to: files/research_notes/{competitor}.md

Step 5: Configure and Run

from pathlib import Path
from claude_agent_framework import create_session
from claude_agent_framework.core.roles import AgentInstanceConfig

agents = [
    AgentInstanceConfig(name="market-researcher", role="worker"),
    AgentInstanceConfig(name="tech-researcher", role="worker"),
    AgentInstanceConfig(name="report-writer", role="synthesizer"),
]

session = create_session(
    "research",
    agent_instances=agents,
    prompts_dir=Path("prompts"),
    template_vars={
        "company_name": "Acme Corp",
        "industry": "Technology",
    },
)

async for msg in session.run("Analyze competitors X, Y, Z"):
    print(msg)

---

Complete Example Structure

my_competitive_intel/
├── main.py                           # Entry point
├── config.yaml                       # Configuration
├── prompts/                          # Business prompts
│   ├── lead_agent.txt                # Coordination strategy
│   ├── researcher.txt                # Research methodology
│   └── report_writer.txt             # Report format
└── .claude/skills/                   # Methodology guidance
    ├── competitive-research/SKILL.md
    └── report-generation/SKILL.md

See examples/production/ for 7 complete, production-ready examples.

---

Two-Layer Prompt Architecture

Prompts are composed from two layers:

Layer	Location	Purpose
Framework	architectures/*/prompts/	Generic role capabilities, workflow rules, dispatching guidelines
Business	Your prompts/ directory	Domain context, Skills references, deliverables, success criteria

Framework Prompt (generic, reusable):

# Role: Research Coordinator

## Core Rules
1. You may ONLY use the Task tool to dispatch sub-agents
2. NEVER perform research, analysis, or writing tasks yourself

## Workflow Phases
1. Plan Phase - Identify parallelizable subtopics
2. Research Phase - Dispatch workers in parallel
3. Processing Phase - Dispatch processor (if configured)
4. Synthesis Phase - Dispatch synthesizer for final output

Business Prompt (domain-specific):

# Competitive Intelligence Coordinator

You are coordinating analysis for ${company_name} in the ${industry} sector.

## Team & Skills
- **Researchers**: Use `competitive-research` Skill
- **Report Writer**: Uses `report-generation` Skill

## Deliverables
- SWOT analysis for each competitor
- Comparative feature matrix
- Strategic recommendations

---

Role-Based Configuration

The framework separates role definitions (architecture-level) from agent instances (business-level):

# Architecture defines role constraints
class ResearchArchitecture(BaseArchitecture):
    def get_role_definitions(self) -> dict[str, RoleDefinition]:
        return {
            "worker": RoleDefinition(
                role_type=RoleType.WORKER,
                required_tools=["WebSearch"],
                cardinality=RoleCardinality.ONE_OR_MORE,  # 1-N workers
                default_model="haiku",
            ),
            "synthesizer": RoleDefinition(
                role_type=RoleType.SYNTHESIZER,
                required_tools=["Write"],
                cardinality=RoleCardinality.EXACTLY_ONE,  # Exactly 1
                default_model="sonnet",
            ),
        }

# Business configures concrete agents
agents = [
    AgentInstanceConfig(name="market-researcher", role="worker"),
    AgentInstanceConfig(name="tech-researcher", role="worker"),
    AgentInstanceConfig(name="report-writer", role="synthesizer", model="opus"),
]

---

Custom Architecture

from claude_agent_framework import register_architecture, BaseArchitecture
from claude_agent_framework.core.roles import RoleDefinition
from claude_agent_framework.core.types import RoleType, RoleCardinality

@register_architecture("my_workflow")
class MyWorkflow(BaseArchitecture):
    name = "my_workflow"
    description = "Custom workflow for my use case"

    def get_role_definitions(self) -> dict[str, RoleDefinition]:
        return {
            "executor": RoleDefinition(
                role_type=RoleType.EXECUTOR,
                description="Execute tasks",
                required_tools=["Read", "Write", "Bash"],
                cardinality=RoleCardinality.ONE_OR_MORE,
            ),
            "reviewer": RoleDefinition(
                role_type=RoleType.CRITIC,
                description="Review results",
                required_tools=["Read"],
                cardinality=RoleCardinality.EXACTLY_ONE,
            ),
        }

    async def execute(self, prompt, tracker=None, transcript=None):
        # Your orchestration logic here
        ...

---

Plugin System

from claude_agent_framework import create_session
from claude_agent_framework.plugins.builtin import (
    MetricsCollectorPlugin,
    CostTrackerPlugin,
)

session = create_session("research")

# Add metrics tracking
metrics = MetricsCollectorPlugin()
session.architecture.add_plugin(metrics)

# Add cost tracking with budget
cost = CostTrackerPlugin(budget_usd=5.0)
session.architecture.add_plugin(cost)

async for msg in session.run("Analyze market"):
    print(msg)

# Get results
print(f"Cost: ${metrics.get_metrics().estimated_cost_usd:.4f}")

---

Session Lifecycle

# Option 1: Manual management
session = create_session("research")
try:
    async for msg in session.run(prompt):
        process(msg)
finally:
    await session.teardown()

# Option 2: Context manager
async with create_session("research") as session:
    results = await session.query(prompt)

---

Output

Each session generates:

logs/session_YYYYMMDD_HHMMSS/
├── transcript.txt      # Human-readable conversation log
└── tool_calls.jsonl    # Structured tool call records

files/
└── <architecture>/     # Architecture-specific outputs

---

Installation

# Basic
pip install claude-agent-framework

# With PDF generation
pip install "claude-agent-framework[pdf]"

# With chart generation
pip install "claude-agent-framework[charts]"

# Full installation
pip install "claude-agent-framework[all]"

---

Documentation

Document	Description
Best Practices	Pattern selection and implementation tips
Role-Based Architecture	Role types, constraints, and agent instantiation
Prompt Writing Guide	Two-layer prompt architecture
Core API Reference	create_session(), AgentSession, BaseArchitecture
Plugin API Reference	Plugin system and lifecycle hooks
Architecture Selection	Choose the right architecture

---

Production Examples

All 7 examples in examples/production/ are complete and production-ready:

Example	Architecture	Business Scenario
01_competitive_intelligence	Research	SaaS competitive analysis
02_pr_code_review	Pipeline	Automated PR review
03_marketing_content	Critic-Actor	Marketing copy optimization
04_it_support	Specialist Pool	IT support routing
05_tech_decision	Debate	Technical decision support
06_code_debugger	Reflexion	Adaptive debugging
07_codebase_analysis	MapReduce	Large codebase analysis

Common Production Patterns

All production examples follow these proven patterns:

Pattern	Implementation
Configuration-Driven	YAML config files with validation, environment-specific settings, no hardcoded values
Structured Results	Consistent JSON schema, programmatic access, multiple formats (JSON/Markdown/PDF)
Robust Error Handling	Try/catch at I/O boundaries, graceful degradation, detailed error context
Comprehensive Logging	JSONL tool tracking, human-readable transcripts, session-based organization
Production Testing	Unit + integration tests, mock external dependencies, 100% pass requirement

When to Use Each Architecture

Architecture	Use When	Example Scenarios
Research	Gathering data from multiple independent sources in parallel	Market research, competitive analysis, literature reviews
Pipeline	Clear sequential stages with quality gates	Code review, content workflows, multi-step approvals
Critic-Actor	Output needs iterative improvement against defined criteria	Content optimization, code refactoring, design iteration
Specialist Pool	Requests need routing to domain experts	Technical support, Q&A systems, diagnostics
Debate	Decision requires balanced pro/con analysis	Technology selection, architecture decisions, vendor evaluation
Reflexion	Problem requires trial-and-error with learning	Debugging, root cause analysis, optimization
MapReduce	Large dataset (500+ files) needs parallel processing	Codebase analysis, batch processing, large-scale audits

---

Requirements

• Python 3.10+
• ANTHROPIC_API_KEY environment variable

License

MIT License - see LICENSE