deepagents-architecture（深度 Agent 架构）

Deep Agents Architecture Decisions

When to Use Deep Agents

Use Deep Agents When You Need:

• Long-horizon tasks - Complex workflows spanning dozens of tool calls
• Planning capabilities - Task decomposition before execution
• Filesystem operations - Reading, writing, and editing files
• Subagent delegation - Isolated task execution with separate context windows
• Persistent memory - Long-term storage across conversations
• Human-in-the-loop - Approval gates for sensitive operations
• Context management - Auto-summarization for long conversations

Consider Alternatives When:

Backend Selection

Backend Comparison

Backend Decision Tree

Need real disk access?
├─ Yes → FilesystemBackend(root_dir="/path")
└─ No
   └─ Need persistence across conversations?
      ├─ Yes → Need mixed ephemeral + persistent?
      │  ├─ Yes → CompositeBackend
      │  └─ No → StoreBackend
      └─ No → StateBackend (default)

CompositeBackend Routing

Route different paths to different storage backends:

from deepagents import create_deep_agent
from deepagents.backends import CompositeBackend, StateBackend, StoreBackend

agent = create_deep_agent(
    backend=CompositeBackend(
        default=StateBackend(),  # Working files (ephemeral)
        routes={
            "/memories/": StoreBackend(store=store),    # Persistent
            "/preferences/": StoreBackend(store=store), # Persistent
        },
    ),
)

Subagent Architecture

When to Use Subagents

Use subagents when:

• Task is complex, multi-step, and can run independently
• Task requires heavy context that would bloat the main thread
• Multiple independent tasks can run in parallel
• You need isolated execution (sandboxing)
• You only care about the final result, not intermediate steps

Don't use subagents when:

• Task is trivial (few tool calls)
• You need to see intermediate reasoning
• Splitting adds latency without benefit
• Task depends on main thread state mid-execution

Subagent Patterns

Pattern 1: Parallel Research

         ┌─────────────┐
         │  Orchestrator│
         └──────┬──────┘
    ┌──────────┼──────────┐
    ▼          ▼          ▼
┌──────┐  ┌──────┐  ┌──────┐
│Task A│  │Task B│  │Task C│
└──┬───┘  └──┬───┘  └──┬───┘
   └──────────┼──────────┘
              ▼
      ┌─────────────┐
      │  Synthesize │
      └─────────────┘

Best for: Research on multiple topics, parallel analysis, batch processing.

Pattern 2: Specialized Agents

research_agent = {
    "name": "researcher",
    "description": "Deep research on complex topics",
    "system_prompt": "You are an expert researcher...",
    "tools": [web_search, document_reader],
}

coder_agent = {
    "name": "coder",
    "description": "Write and review code",
    "system_prompt": "You are an expert programmer...",
    "tools": [code_executor, linter],
}

agent = create_deep_agent(subagents=[research_agent, coder_agent])

Best for: Domain-specific expertise, different tool sets per task type.

Pattern 3: Pre-compiled Subagents

from deepagents import CompiledSubAgent, create_deep_agent

# Use existing LangGraph graph as subagent
custom_graph = create_react_agent(model=..., tools=...)

agent = create_deep_agent(
    subagents=[CompiledSubAgent(
        name="custom-workflow",
        description="Runs specialized workflow",
        runnable=custom_graph
    )]
)

Best for: Reusing existing LangGraph graphs, complex custom workflows.

Middleware Architecture

Built-in Middleware Stack

Deep Agents applies middleware in this order:

1. TodoListMiddleware - Task planning with write_todos/read_todos
2. FilesystemMiddleware - File ops: ls, read_file, write_file, edit_file, glob, grep, execute
3. SubAgentMiddleware - Delegation via task tool
4. SummarizationMiddleware - Auto-summarizes at ~85% context or 170k tokens
5. AnthropicPromptCachingMiddleware - Caches system prompts (Anthropic only)
6. PatchToolCallsMiddleware - Fixes dangling tool calls from interruptions
7. HumanInTheLoopMiddleware - Pauses for approval (if interrupt_on configured)

Custom Middleware Placement

from langchain.agents.middleware import AgentMiddleware

class MyMiddleware(AgentMiddleware):
    tools = [my_custom_tool]

    def transform_request(self, request):
        # Modify system prompt, inject context
        return request

    def transform_response(self, response):
        # Post-process, log, filter
        return response

# Custom middleware added AFTER built-in stack
agent = create_deep_agent(middleware=[MyMiddleware()])

Middleware vs Tools Decision

Subagent Middleware Inheritance

Subagents receive their own middleware stack by default:

• TodoListMiddleware
• FilesystemMiddleware (shared backend)
• SummarizationMiddleware
• AnthropicPromptCachingMiddleware
• PatchToolCallsMiddleware

Override with default_middleware=[] in SubAgentMiddleware or per-subagent middleware key.

Gates: architecture decisions before implementation

Complete in order. A step passes only when the stated artifact exists in the design note, ADR stub, or ticket; internal intent alone does not count.

1. Fit - Confirm Deep Agents vs alternatives (see tables above).

- Pass: Short written rationale that either names one matching "Use Deep Agents When You Need" bullet or one "Consider Alternatives" row plus the chosen alternative.

1. Backend - Match the Backend Decision Tree to a concrete choice.

- Pass: Backend name(s) from the Backend Comparison table; if FilesystemBackend or CompositeBackend, root_dir and any route prefixes are written down (path placeholders OK).

1. Subagents - Decide delegation boundaries.

- Pass: Either "no subagents" plus one sentence why or a named list where each subagent maps to at least one "When to Use Subagents" reason; parallel plans state what merges outputs.

1. Human-in-the-loop - Approval surface.

- Pass: Explicit list of tools/operations that use interrupt_on, or "no HITL" plus one-line risk acceptance.

1. Middleware - Custom vs built-in only.

- Pass: Either "custom middleware: none" or each custom piece named, placed after the built-in stack, and tied to prompt injection, tools, or request/response transforms.

1. Context - Long threads and large inputs.

- Pass: Stated plan for default summarization behavior (~85% context / ~170k tokens) or an alternative cap; large files handled via references/chunking or equivalent, named in text.

1. Checkpointing - Resume and durability.

- Pass: Checkpoint/checkpointer approach named for the graph or "none" with one-line rationale (e.g. ephemeral demo only).