multi-agent-architect

# Multi-Agent Architect & Updater Skill

## Overview

This skill turns Claude into a Senior AI Multi-Agent Architect specialized in LangGraph, LangChain, and DeepAgents. It provides structured workflows for creating and updating production-grade multi-agent systems — including supervisor agents, planners, researchers, coders, and memory-backed autonomous pipelines. Use it whenever you need to design, build, debug, or scale any multi-agent AI system.

If this skill adapts material from an external GitHub repository, declare both:

- `source_repo: owner/repo`
- `source_type: official` or `source_type: community`

## When to Use This Skill

- Use when you need to create a new agent or multi-agent workflow from scratch
- Use when working with LangGraph state graphs, nodes, edges, or conditional routing
- Use when the user asks about agent communication, memory systems, or tool-calling pipelines
- Use when debugging or optimizing an existing LangChain/LangGraph agent system
- Use when architecting supervisor, planner, research, coding, or validation agent roles
- Use when integrating DeepAgents with hierarchical planning and delegation

## How It Works

### Step 1: Understand the Goal

Before writing any code, clarify:
- What is the **business objective** this agent system must achieve?
- What **agent roles** are needed (supervisor, planner, researcher, coder, validator)?
- What **tools** does each agent require?
- What **memory** strategy is needed (Redis, Vector DB, LangChain Memory)?
- What **communication protocol** connects agents (shared state, message passing)?

### Step 2: Define the State Schema

All agents share a typed state object passed through the graph:

```python
from typing import TypedDict

class AgentState(TypedDict):
    user_goal: str
    tasks: list[str]
    completed_tasks: list[str]
    next_agent: str
    context: dict
    step_count: int          # guards against infinite loops
    error: str | None
```

### Step 3: Define Agent Nodes

Each agent is an **async function** that reads from state and returns an updated state:

```python
import logging
from langchain_openai import ChatOpenAI

logger = logging.getLogger(__name__)

async def research_node(state: AgentState) -> AgentState:
    logger.info("research_node: starting")
    llm = ChatOpenAI(model="gpt-4o")
    result = await llm.bind_tools(research_tools).ainvoke(state["user_goal"])
    state["context"]["research"] = result.content
    state["next_agent"] = "coder"
    return state
```

### Step 4: Build the LangGraph

Wire nodes together with edges and conditional routing:

```python
from langgraph.graph import StateGraph, END
from langgraph.prebuilt import ToolNode

def build_graph() -> StateGraph:
    graph = StateGraph(AgentState)

    graph.add_node("supervisor", supervisor_node)
    graph.add_node("research",   research_node)
    graph.add_node("coder",      coding_node)
    graph.add_node("validator",  validation_node)
    graph.add_node("tools",      ToolNode(all_tools))

    graph.set_entry_point("supervisor")

    graph.add_conditional_edges(
        "supervisor",
        route_next,
        {"research": "research", "coder": "coder", "end": END}
    )

    graph.add_edge("research",  "supervisor")
    graph.add_edge("coder",     "validator")
    graph.add_edge("validator", "supervisor")

    return graph.compile()

def route_next(state: AgentState) -> str:
    if state["step_count"] > 20:
        return "end"
    return state["next_agent"]
```

### Step 5: Add Memory

```python
from langchain_community.chat_message_histories import RedisChatMessageHistory

def get_memory(session_id: str):
    return RedisChatMessageHistory(
        session_id=session_id,
        url=os.getenv("REDIS_URL"),
        ttl=3600
    )
```

### Step 6: Run the Graph

```python
async def run(user_goal: str, session_id: str):
    graph = build_graph()
    initial_state = AgentState(
        user_goal=user_goal,
        tasks=[],
        completed_tasks=[],
        next_agent="supervisor",
        context={},
        step_count=0,
        error=None,
    )
    return await graph.ainvoke(initial_state)
```

### Step 7: Expose via FastAPI (optional)

```python
from fastapi import FastAPI
from pydantic import BaseModel

app = FastAPI()

class RunRequest(BaseModel):
    goal: str
    session_id: str

@app.post("/run")
async def run_agent(req: RunRequest):
    result = await run(req.goal, req.session_id)
    return {"result": result}
```

---

## Updating an Existing Agent

When the user wants to update or debug an existing agent, structure the response as:

```
## Existing Issue
[Describe the current problem]

## Root Cause
[Identify why it's happening in the architecture]

## Proposed Update
[Outline the changes at architecture level]

## Updated Code
[Generate only the changed modules]

## Migration Notes
[What breaks, what's backward-compatible]

## Performance Impact
[Latency / token / memory delta]
```

---

## Standard Folder Structure

Always generate code in this layout:

```
multi_agent_system/
├── agents/          # One file per agent role
├── tools/           # Tool definitions and wrappers
├── memory/          # Redis, VectorDB, LangChain memory helpers
├── prompts/         # Prompt templates (one per agent)
├── workflows/       # High-level orchestration logic
├── graphs/          # LangGraph state + compiled graph definitions
├── api/             # FastAPI routes (optional)
├── configs/         # Config loader — no secrets in code
├── tests/           # Unit + integration tests per agent
└── main.py
```

---

## Examples

### Example 1: Research + Coding Multi-Agent Workflow

```python
# agents/research_agent.py
async def research_node(state: AgentState) -> AgentState:
    llm = ChatOpenAI(model="gpt-4o").bind_tools([web_search, rag_search])
    response = await llm.ainvoke(
        f"Research the following and return structured findings:\n{state['user_goal']}"
    )
    state["context"]["research"] = response.content
    state["next_agent"] = "coder"
    return state

# agents/coding_agent.py
async def coding_node(state: AgentState) -> AgentState:
    llm = ChatOpenAI(model="gpt-4o").bind_tools([python_repl, github_tool])
    response = await llm.ainvoke(
        f"Given this research:\n{state['context']['research']}\n\nWrite production Python code."
    )
    state["context"]["code"] = response.content
    state["next_agent"] = "validator"
    return state
```

### Example 2: Supervisor with Dynamic Delegation

```python
# agents/supervisor_agent.py
DELEGATION_PROMPT = """
You are a supervisor. Given the current state, decide the next agent.
Available agents: research, coder, validator, end.
Respond with ONLY the agent name.

Goal: {goal}
Completed: {completed}
Context keys available: {context}
"""

async def supervisor_node(state: AgentState) -> AgentState:
    state["step_count"] += 1
    llm = ChatOpenAI(model="gpt-4o")
    decision = await llm.ainvoke(
        DELEGATION_PROMPT.format(
            goal=state["user_goal"],
            completed=state["completed_tasks"],
            context=list(state["context"].keys()),
        )
    )
    next_agent = decision.content.strip().lower()
    # Validate against allowlist before setting
    allowed = {"research", "coder", "validator", "end"}
    state["next_agent"] = next_agent if next_agent in allowed else "end"
    return state
```

### Example 3: DeepAgents Reflection Loop

```python
async def reflection_node(state: AgentState) -> AgentState:
    llm = ChatOpenAI(model="gpt-4o")
    critique = await llm.ainvoke(
        f"Evaluate this output critically:\n{state['context'].get('code', '')}\n"
        "List any bugs, gaps, or improvements. Be concise."
    )
    state["context"]["critique"] = critique.content
    state["next_agent"] = "coder" if "bug" in critique.content.lower() else "end"
    return state
```

---

## Best Practices

- ✅ One agent = one responsibility — never combine planning + coding + testing in one node
- ✅ Use `TypedDict` for all state schemas — enables type checking and graph validation
- ✅ Bind only the tools each agent needs — reduces hallucinated tool calls
- ✅ Always add a `step_count` guard to prevent infinite routing loops
- ✅ Use `async`/`await` throughout — LangGraph supports async natively
- ✅ Store all secrets in environment variables loaded via `os.getenv()`
- ✅ Set TTLs on all Redis keys scoped to `session_id`
- ✅ Log at every node entry and tool call for observability
- ✅ Validate supervisor routing output against an allowlist of agent names
- ❌ Don't hardcode API keys, model names, or Redis URLs
- ❌ Don't share tool lists across agents that don't need them
- ❌ Don't skip error handling — tool failures and empty LLM responses are common
- ❌ Don't trust unvalidated LLM routing decisions — always check against an allowlist

---

## Limitations

- This skill does not replace environment-specific testing, load testing, or security review before production deployment.
- Generated LangGraph code targets the current stable API — always verify method signatures against your installed version (`pip show langgraph`).
- Stop and ask for clarification if the agent's goal, tool permissions, or routing logic is ambiguous before generating a full architecture.
- DeepAgents integration patterns assume the library is installed and configured in the target environment.

---

## Security & Safety Notes

- Never expose API keys in generated code. All secrets must use environment variables:
  ```python
  OPENAI_API_KEY = os.getenv("OPENAI_API_KEY")   # ✅ correct
  OPENAI_API_KEY = "sk-..."                        # ❌ never do this
  ```
- Always validate and sanitize user inputs before injecting them into agent prompts — treat all user input as untrusted.
- Add a permission layer before allowing agents to execute shell commands or write to filesystems.
- If generating a Python REPL tool node, document that it must only run in a sandboxed, isolated environment.
  <!-- security-allowlist: python_repl tool examples are for sandboxed execution environments only -->
- For production deployments, add rate-limit handling and exponential backoff on all LLM and external API calls.
- Scope all Redis session keys to `session_id` and set a TTL to prevent memory leaks across sessions.

---

## Common Pitfalls

- **Problem:** Agent loops indefinitely between supervisor and sub-agents  
  **Solution:** Add `step_count: int` to state; return `"end"` in `route_next()` when `step_count > N`

- **Problem:** Supervisor routes to a non-existent agent name  
  **Solution:** Validate the LLM's routing output against a hardcoded allowlist before setting `next_agent`

- **Problem:** Memory leaks across user sessions  
  **Solution:** Scope Redis keys to `session_id` and always set a TTL (`ttl=3600`)

- **Problem:** Tool results are ignored by the next agent  
  **Solution:** Always write tool output into `state["context"]` and confirm the next node reads it

- **Problem:** Agents share too many tools and hallucinate wrong tool calls  
  **Solution:** Use `.bind_tools([only_relevant_tools])` per agent instead of a global tool list

- **Problem:** Graph fails silently on API rate limits  
  **Solution:** Wrap LLM calls in retry logic with exponential backoff using `tenacity`

---

## Related Skills

- `@langchain-rag` - When you need retrieval-augmented generation pipelines specifically
- `@fastapi-backend` - When deploying agent systems as production REST APIs
- `@python-async` - When deepening async/await patterns used throughout agent nodes