Multi Agent

简体中文 • 日本語 • 한국어 • Español • Deutsch • Dansk

multi-agent

The definitive catalog of AI agent patterns. One definition, any framework.

Quick Start • Agent Catalog • Smart Enhancements • Patterns • Playground • Docs • Contributing

---

48 catalog agent definitions. 11 categories. 8 orchestration patterns. 6 framework adapters. 8 export targets. Zero lock-in.

multi-agent is a framework-agnostic catalog of production-ready AI agent patterns. Define your agents once in YAML, run them on CrewAI, LangGraph, OpenAI Agents SDK, Claude SDK, Google ADK, or smolagents.

Stop reinventing agents. Start composing them.

Why multi-agent?

| | multi-agent | CrewAI | LangGraph | OpenAI SDK | Claude SDK | |---|:---:|:---:|:---:|:---:|:---:| | Framework-agnostic definitions | Yes | No | No | No | No | | Export to any AI platform | Yes | No | No | No | No | | Reusable agent catalog | Yes | Framework-specific | Framework-specific | Framework-specific | Framework-specific | | Pattern library (8 patterns) | Yes | 2 | 3 | 2 | 2 | | Built-in cost estimation | Yes | No | No | No | No | | Agent recommendation engine | Yes | No | No | No | No | | Works with any LLM | Yes | Yes | Yes | OpenAI only | Claude only | | MCP native | Yes | Partial | Adapter | Yes | Yes |

Quick Start

pip install multi-agent

1. Browse the catalog

multiagent search "code review"

Found 3 agents matching "code review":

  code/code-reviewer     Review PRs for bugs, style, and security
  code/test-writer       Generate tests for changed code
  code/refactorer        Suggest and apply refactoring improvements

Recommended pattern: supervisor-worker (1 reviewer + N specialists)
Estimated cost: ~$0.03/review (Claude Haiku) to ~$0.25/review (GPT-4o)

2. Use an agent definition

from multiagent import Catalog, patterns

# Load agents from the catalog
catalog = Catalog()
reviewer = catalog.load("code/code-reviewer")
test_writer = catalog.load("code/test-writer")

# Describe a portable pattern; execution belongs to a framework adapter
team = patterns.supervisor_worker(
    supervisor=reviewer,
    workers=[test_writer],
    model="claude-sonnet-4-6"  # or any model
)
print(team.describe())

3. Or use with your favorite framework

# Build a framework-neutral OpenAI agent-as-tool plan without importing its SDK
from multiagent.adapters import openai_sdk
plan = openai_sdk.to_agent_tool_config(reviewer, [test_writer])

# Optional runtime adapters create native objects when their extras are installed
# pip install "multi-agent[openai]"
agent = openai_sdk.from_catalog("code/code-reviewer")
# from agents import Runner
# result = Runner.run_sync(agent, "Review this code")

4. Export to any AI platform

# Export for Claude Code (.claude/agents subagents)
multiagent export code/code-reviewer claude-code -o .claude/agents

# Export as a portable AgentSkills-style SKILL.md file
multiagent export code/code-reviewer agentskill -o .agents/skills/code-reviewer

# Export as an A2A Agent Card JSON document
multiagent export code/code-reviewer a2a-agent-card -o ./agent-cards

# Export for Codex / OpenClaw (AGENTS.md format)
multiagent export code/code-reviewer codex

# Export for Codex project config (.codex/config.toml snippet)
mkdir -p .codex
multiagent export code/code-reviewer codex-config > .codex/config.toml

# Export for Google Gemini / ADK
multiagent export code/code-reviewer gemini -o ./adk-agents

# Export for ChatGPT (Custom GPT instructions)
multiagent export code/code-reviewer chatgpt

# Export just the system prompt (works with ANY LLM)
multiagent export code/code-reviewer raw

# Bulk export all agents in a category
multiagent export-all claude-code -o .claude/agents -c code

| Target | Format | Works With | |--------|--------|------------| | claude-code | .md subagent files | Claude Code .claude/agents/ | | agentskill | SKILL.md-style Markdown | AgentSkills-compatible tools | | a2a-agent-card | Agent Card JSON | A2A discovery via .well-known/agent-card.json | | codex | AGENTS.md sections | OpenAI Codex, OpenClaw | | codex-config | .codex/config.toml snippet | OpenAI Codex multi-agent roles | | gemini | ADK YAML config | Google Gemini, Vertex AI | | chatgpt | System instructions | ChatGPT, Custom GPTs | | raw | Plain system prompt | Any LLM — Ollama, LM Studio, llama.cpp, vLLM, etc. |

Route before exporting

# Dry-run agent selection and include export commands for the chosen target
multiagent route "review this PR and write missing tests" --target a2a-agent-card

# Machine-readable route decision with target export plan
multiagent route "review this PR and write missing tests" --target codex-config --json

# Regression-test the built-in routing corpus
multiagent eval-routing --json

# CI-friendly score gates; target hints may evolve faster than agent matching
multiagent eval-routing \
  --min-agent-score 1.0 \
  --min-pattern-score 1.0 \
  --min-target-score 0.95 \
  --min-forbidden-score 1.0 \
  --min-risk-score 1.0 \
  --min-context-score 1.0 \
  --min-policy-score 1.0

# Validate all schemas, slugs, duplicate names, and cross-agent references
multiagent validate

# Route directly into a framework-native plan instead of a file export
multiagent route "review this PR and write missing tests" --target openai-agents --json
multiagent route "research three sources with Google ADK" --target adk --json

Route JSON includes risk, context, and policy blocks. The policy makes control mode, bounded fan-out, delegation contracts, trust boundaries, approval requirements, and stop conditions explicit. This command is a dry run: it never executes an agent or contacts a model provider. See [08 -- Human Review Gates](cookbook/08-human-review-gates.md) and [09 -- Safe Delegation Policy](cookbook/09-safe-delegation-policy.md).

Agent Catalog

Every agent is defined in a simple, readable YAML format:

# src/multiagent/catalog_data/code/code-reviewer.yaml
name: code-reviewer
version: "1.0"
description: Reviews code changes for bugs, security issues, and style violations
category: code

system_prompt: |
  You are an expert code reviewer. Analyze the provided code changes and identify:
  1. Bugs and logic errors
  2. Security vulnerabilities (OWASP Top 10)
  3. Performance issues
  4. Style and readability improvements
  
  Be specific. Reference line numbers. Suggest fixes.

tools:
  - type: mcp
    server: filesystem
  - type: function
    name: search_codebase
    description: Search for related code in the repository

parameters:
  temperature: 0.1
  max_tokens: 4096

cost_profile:
  tokens_per_run: ~2000
  recommended_model: claude-haiku-4-5  # Best cost/quality for reviews
  estimated_cost_usd: 0.003

works_with:
  - code/test-writer      # Generate tests for flagged code
  - code/refactorer        # Apply suggested improvements
  
recommended_patterns:
  - supervisor-worker      # Reviewer supervises specialist agents
  - sequential             # Review → Test → Refactor pipeline

# Optional schema v2 metadata for routing, safety, observability, and protocols
orchestration:
  control_mode: router
  execution_mode: dry_run
safety:
  side_effect_risk: low
  requires_human_review: true
observability:
  trace_tags: [code-review]
  eval_criteria: [finds-bugs, suggests-fixes]
outputs:
  expected_artifacts: [review-comments]
context:
  loading: trigger
  max_context_tokens: 4096
protocols:
  a2a:
    expose: true

Full Catalog

| Category | Agents | Description | |----------|--------|-------------| | [code/](src/multiagent/catalogdata/code/) | code-reviewer code-generator test-writer refactorer debugger security-auditor documentation-writer pr-summarizer | Software development lifecycle | | [research/](src/multiagent/catalogdata/research/) | deep-researcher web-scraper fact-checker paper-analyst competitive-intel | Research and analysis | | [data/](src/multiagent/catalogdata/data/) | data-analyst sql-generator report-writer | Data engineering and analysis | | [devops/](src/multiagent/catalogdata/devops/) | ci-cd-agent infra-provisioner monitoring-agent incident-responder | Infrastructure and operations | | [content/](src/multiagent/catalogdata/content/) | writer editor translator seo-optimizer | Content creation pipeline | | [finance/](src/multiagent/catalogdata/finance/) | trading-analyst portfolio-optimizer financial-reporter fraud-detector tax-advisor | Financial analysis and compliance | | [support/](src/multiagent/catalogdata/support/) | customer-support ticket-router knowledge-base-builder escalation-agent | Customer service pipeline | | [legal/](src/multiagent/catalogdata/legal/) | contract-reviewer legal-researcher compliance-checker document-drafter | Legal and compliance | | [personal/](src/multiagent/catalogdata/personal/) | email-assistant meeting-scheduler note-taker task-manager | Personal productivity | | [security/](src/multiagent/catalogdata/security/) | vulnerability-scanner log-analyzer access-reviewer incident-analyst | Security operations | | [orchestration/](src/multiagent/catalog_data/orchestration/) | task-router cost-optimizer quality-gate | Meta-agents for coordination |

Patterns

Eight battle-tested orchestration patterns, each with runnable examples:

Pattern Overview

                    ┌─────────────┐
                    │  Supervisor  │ ── Pattern 1: Supervisor/Worker
                    └──────┬──────┘    Central agent delegates to specialists
                     ┌─────┼─────┐
                     ▼     ▼     ▼
                   [W1]  [W2]  [W3]

    [A] → [B] → [C]                ── Pattern 2: Sequential Pipeline
                                       Linear chain of specialized agents

    ┌──→ [A] ──┐
    │          │
    ├──→ [B] ──┼──→ [Merge]        ── Pattern 3: Parallel Fan-Out
    │          │                       Independent tasks run concurrently
    └──→ [C] ──┘

    [A] ←──→ [B]                   ── Pattern 4: Reflection/Loop
                                       Iterative refinement between agents

    [A] ──handoff──→ [B] ──handoff──→ [C]  ── Pattern 5: Handoff
                                              Agent transfers full control

    ┌─[A]─┐                        ── Pattern 6: Group Chat
    │ [B] │ ← Selector                Shared conversation, dynamic speaker
    └─[C]─┘

    [A]─┬─[B]                      ── Pattern 7: DAG (Directed Acyclic Graph)
        └─[C]──[D]                    Conditional branching and merging

    ┌Worktree1: [A]┐               ── Pattern 8: Split-and-Merge
    │Worktree2: [B]│→ Git Merge       Isolated parallel work, merged at end
    └Worktree3: [C]┘

| Pattern | When to Use | Complexity | Latency | Example | |---------|------------|:----------:|:-------:|---------| | [Supervisor/Worker](docs/patterns/supervisor-worker.md) | Complex tasks with clear subtasks | Medium | Medium | Code review team | | [Sequential](docs/patterns/sequential.md) | Step-by-step processing | Low | High | Content pipeline | | [Parallel](docs/patterns/parallel.md) | Independent tasks | Low | Low | Multi-source research | | [Reflection](docs/patterns/reflection.md) | Quality-critical output | Medium | Medium | Legal document drafting | | [Handoff](docs/patterns/handoff.md) | Escalation and routing | Low | Low | Customer support tiers | | [Group Chat](docs/patterns/group-chat.md) | Brainstorming, debate | High | High | Design review | | [DAG](docs/patterns/dag.md) | Complex conditional workflows | High | Variable | CI/CD pipeline | | [Split-and-Merge](docs/patterns/split-and-merge.md) | Large parallel code changes | Medium | Low | Multi-file refactoring |

Frameworks

multi-agent provides adapter modules for the main Python agent frameworks:

| Framework | Adapter | Current scope | |-----------|---------|---------------| | [CrewAI](docs/frameworks/comparison.md) | multiagent.adapters.crewai | Agent/Crew conversion; Flow template config | | [LangGraph](docs/frameworks/comparison.md) | multiagent.adapters.langgraph | Node/flow config conversion | | [OpenAI Agents SDK](docs/frameworks/comparison.md) | multiagent.adapters.openai_sdk | Agent conversion; handoff and agent-as-tool plans | | [Claude Agent SDK](docs/frameworks/comparison.md) | multiagent.adapters.claude_sdk | Message/subagent config conversion | | [Google ADK](docs/frameworks/comparison.md) | multiagent.adapters.google_adk | ADK config conversion; sequential/parallel workflow plans | | [smolagents](docs/frameworks/comparison.md) | multiagent.adapters.smolagents | Agent config conversion; manager/managed-agent plans |

Adapter template helpers return plain dictionaries, so they can be inspected, tested, or translated into framework code before any framework package is installed:

from multiagent import Catalog
from multiagent.adapters import crewai, google_adk, openai_sdk, smolagents

catalog = Catalog()
reviewer = catalog.load("code/code-reviewer")
test_writer = catalog.load("code/test-writer")

handoff_plan = openai_sdk.to_handoff_config(reviewer, [test_writer])
tool_plan = openai_sdk.to_agent_tool_config(reviewer, [test_writer])
adk_parallel = google_adk.to_workflow_config([reviewer, test_writer], workflow="parallel")
flow = crewai.to_flow_config([reviewer, test_writer], flow_name="CodeReviewFlow", human_feedback=True)
manager = smolagents.to_manager_config(reviewer, [test_writer])

The router can also emit framework-native dry-run plans directly:

multiagent route "review this PR and write missing tests with OpenAI Agents SDK" \
  --target openai-agents \
  --json

multiagent route "research three sources with Google ADK" --target adk --json
multiagent route "build a deterministic CrewAI review flow" --target crewai-flow --json
multiagent route "coordinate workers with smolagents" --target smolagents-manager --json

Don't see your framework? [Submit an adapter](CONTRIBUTING.md#adding-an-adapter).

Protocol Support

Built for the 2026 protocol stack:

| Protocol | Purpose | Support | |----------|---------|:-------:| | [MCP](docs/protocols/mcp.md) (Model Context Protocol) | Agent ↔ Tool | Native | | [A2A](docs/protocols/a2a.md) (Agent-to-Agent) | Agent ↔ Agent | Native | | [AG-UI](docs/protocols/ag-ui.md) (Agent-to-UI) | Agent ↔ Frontend | Planned |

Smart Enhancements

Make any agent smarter with research-backed prompt engineering techniques:

# Enhance with category-tuned profile
multiagent enhance code/code-reviewer

# Apply all 8 techniques
multiagent enhance code/code-reviewer -p all

# Enhance + export to Claude Code
multiagent enhance code/code-reviewer -p all -t claude-code -o .claude/agents

| Enhancement | Effect | Source | |------------|--------|--------| | reasoning | +20% task completion | OpenAI SWE-bench | | error_recovery | 5-level retry hierarchy | Anthropic engineering | | verification | Self-check before output | Claude Code internal | | confidence | -40-60% hallucination | Academic research | | tool_discipline | Faster, fewer errors | OpenAI GPT-5.4 guide | | failure_modes | Avoids 6 anti-patterns | 120+ leaked prompts study | | context_management | Better long-running tasks | LangChain context engineering | | information_priority | Facts over guessing | Manus AI / Anthropic |

from multiagent import Catalog, enhance_agent

catalog = Catalog()
agent = catalog.load("code/code-reviewer")
smart_agent = enhance_agent(agent, profile="all")  # All 8 techniques applied

Agent Composition Visualizer

Auto-generate Mermaid diagrams for agent teams:

multiagent visualize code/code-reviewer code

…

## Source & license

This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.

- **Author:** [Hovborg](https://github.com/Hovborg)
- **Source:** [Hovborg/multi-agent](https://github.com/Hovborg/multi-agent)
- **License:** MIT

Install and usage instructions live in the source repository linked above.