Deep Agents Planning and Todos

Master the write_todos tool for effective task planning and decomposition in Deep Agents.

Use This Skill When

You need to break down complex multi-step tasks (3+ steps) into trackable subtasks.
You want to show users the plan before executing (user approval workflow).
You're debugging why todos aren't completing as expected.
You need patterns for different task types (research, coding, analysis, document processing).
You want to visualize todo progression from LangSmith traces.

When To Use write_todos

| Use write_todos | Execute Directly | |-----------------|------------------| | ✅ Complex multi-step tasks (3-6 steps) | ✅ Simple 1-2 step queries | | ✅ Tasks requiring user approval first | ✅ Single tool calls | | ✅ Long-running workflows needing progress tracking | ✅ Quick information lookups | | ✅ Tasks where planning adds clarity | ✅ Straightforward API calls |

Decision rule: If you'd benefit from showing the user "Here's my plan..." before starting, use write_todos.

Quick Start

from deepagents import create_deep_agent

# TodoListMiddleware is included by default in create_deep_agent
agent = create_deep_agent(
    model="anthropic:claude-sonnet-4-5-20250929",
    tools=[search_tool, summarize_tool],
    system_prompt="You are a research assistant. Use write_todos for multi-step tasks."
)

# Agent workflow:
# 1. Call write_todos with initial plan
# 2. Ask user: "Does this plan look good?"
# 3. User approves → start executing
# 4. Update the todo list as work progresses
# 5. Keep todos aligned with the current plan and execution state

Example todo creation:

# Agent calls write_todos internally:
{
  "name": "write_todos",
  "arguments": {
    "todos": [
      {"content": "Search for papers on LLM agents", "status": "pending"},
      {"content": "Read and extract findings from top 5 papers", "status": "pending"},
      {"content": "Identify common themes", "status": "pending"},
      {"content": "Write summary report", "status": "pending"}
    ]
  }
}

Todo Structure and API

Two-Field Structure

{
  "content": "Task description (clear, actionable)",
  "status": "pending" | "in_progress" | "completed"
}

Key Constraints

Full-list updates: Treat each write_todos call as a full state update and include all active todos.

Per-turn discipline: Prefer one write_todos update per model turn to avoid conflicting plan changes.

Best granularity: Keep lists to 3-6 items maximum (avoid over-fragmentation).

Tooling Note

Deep Agents documentation describes write_todos as the built-in interface for todo planning/tracking. Keep todo state accurate by rewriting the list with updated statuses as execution progresses.

Status Lifecycle

pending → in_progress → completed

Best practices:

Create todos with "status": "pending" for newly planned work.
Update to "in_progress" when starting work on a todo.
Mark "completed" when finished (don't delete - keeps context).
For interactive workflows, ask user approval ("Does this plan look good?") before starting execution.

Typical workflow:

# Step 1: Create initial plan (all pending)
write_todos([
  {"content": "Research topic", "status": "pending"},
  {"content": "Write summary", "status": "pending"}
])

# Step 2: Ask user approval
# User: "Yes, proceed"

# Step 3: Start first task
write_todos([
  {"content": "Research topic", "status": "in_progress"},
  {"content": "Write summary", "status": "pending"}
])

# Step 4: Complete first task, start second
write_todos([
  {"content": "Research topic", "status": "completed"},
  {"content": "Write summary", "status": "in_progress"}
])

# Step 5: Finish all tasks
write_todos([
  {"content": "Research topic", "status": "completed"},
  {"content": "Write summary", "status": "completed"}
])

Todo Patterns By Task Type

Quick Reference

| Task Type | Pattern | Example Todos | |-----------|---------|---------------| | Research | gather → synthesize → report | Search docs, Read examples, Analyze patterns, Synthesize findings | | Coding | design → implement → test | Design API, Implement endpoints, Write tests, Test end-to-end | | Analysis | collect → process → analyze | Collect data, Process traces, Analyze patterns, Visualize results | | Document Processing | read → extract → transform | Read files, Extract key info, Transform format, Output result |

For detailed patterns with code examples, see references/todo-patterns.md.

Best Practices

✅ DO

Granularity: Keep lists to 3-6 items max (clear milestones, not micro-tasks).
Naming: Use clear, action-oriented descriptions ("Search for X", "Analyze Y").
User interaction: Always ask approval before executing plan.
Status updates: Update promptly as items complete (don't skip status transitions).
Context management: Use with filesystem tools for complex workflows.

⚠️ DON'T

Over-fragment: Avoid 10+ todos (too granular, hard to track).
Vague descriptions: "Do research" → "Search LangChain docs for Deep Agents overview".
Skip approval: Don't start executing without user confirmation.
Forget updates: Always update status when transitioning tasks.
Drop existing context: Include existing active items when rewriting todos.

Troubleshooting

Todo Not Completing

Symptom: Todo stuck in in_progress, agent loops or gets confused.

Causes & fixes:

Missing status update → Ensure agent updates status when task finishes.
Unclear completion criteria → Make content more specific ("Read 5 papers" vs "Do research").
Agent forgot about todos → Add to system prompt: "Use write_todos to maintain and update the plan as work progresses."

Agent Ignoring Todos

Symptom: Agent creates todos but doesn't follow them.

Causes & fixes:

Missing system prompt guidance → Add: "Follow the todo list. Update status as you complete each item."
One-off task (doesn't need todos) → Use direct execution for simple queries.
Conflicting instructions → Remove competing planning instructions from prompt.

Too Many Todos

Symptom: 10+ todos, hard to track, agent overwhelmed.

Causes & fixes:

Over-planning → Simplify to 3-6 high-level milestones.
Nested subtasks → Use todo hierarchy pattern (see references/todo-patterns.md).
Wrong abstraction → Consider breaking into multiple agent invocations.

Lost Context

Symptom: Agent loses track of what's been done.

Causes & fixes:

No filesystem persistence → Use FilesystemBackend or StoreBackend for long sessions.
Not maintaining todos → Update write_todos whenever status changes or scope shifts.
Memory issues → Use MemoryMiddleware for long-term context.

Visualizing Todos

Use the included script to parse LangSmith traces and visualize todo progression:

# Export trace from LangSmith (download JSON)
# Then run:
uv run skills/deepagents-planning-todos/scripts/visualize_todos.py trace.json

# Show Mermaid diagram:
uv run skills/deepagents-planning-todos/scripts/visualize_todos.py trace.json --format mermaid

# Show full timeline:
uv run skills/deepagents-planning-todos/scripts/visualize_todos.py trace.json --show-timeline

Output example:

Todo Timeline for trace abc123:

Initial Plan (Step 1):
  ⏳ [pending] Search for papers on LLM agents
  ⏳ [pending] Read and extract findings
  ⏳ [pending] Identify common themes
  ⏳ [pending] Write summary report

Final State:
  ✅ [completed] Search for papers on LLM agents
  ✅ [completed] Read and extract findings
  ✅ [completed] Identify common themes
  ✅ [completed] Write summary report

Resources

References (detailed patterns):

references/todo-patterns.md: Task-specific patterns with code examples

Examples (working code):

assets/examples/todo-driven-agent/: Research agent demonstrating full workflow

Example structures (templates):

assets/todo-structures/research-todos.json: Research task breakdown
assets/todo-structures/coding-todos.json: Coding task breakdown

External docs:

Deep Agents overview: https://docs.langchain.com/oss/python/deepagents/overview
Deep Agents customization (middleware defaults): https://docs.langchain.com/oss/python/deepagents/customization
Deep Agents harness (planning capabilities): https://docs.langchain.com/oss/python/deepagents/harness
LangChain To-do middleware: https://docs.langchain.com/oss/python/langchain/middleware/built-in
LangSmith tracing for Deep Agents: https://docs.langchain.com/langsmith/trace-deep-agents