[FEATURE] Hierarchical planning and execution.

[ANogin]

Preflight Checklist

• I have searched existing requests and this feature hasn't been requested yet
• This is a single feature request (not multiple features)

Problem Statement

When implementing something sufficiently complicated, Claude would often create a good plan, but then be unable to execute because the steps of the plan are not something it can 1-shot correctly. It would be great to be able to put into a "hierarchical planning and execution" mode, where for each step of the outer plan, it would go into an inner planning more, then execute the inter plan (with or without my approval), then proceed onto the next outer step. Of course both the inner and outer plan should have a Todo.

Proposed Solution

A natural way would be for Claude code to know to spawn a effectively a copy of itself as an agent for executing each step of the outer plan - of course providing that other copy with all the relevant context / etc

Alternative Solutions

No response

Priority

High - Significant impact on productivity

Feature Category

Interactive mode (TUI)

Use Case Example

No response

Additional Context

No response

[github-actions]

---

Found 2 possible duplicate issues:

1. Feature Request: A Structured Workflow for Complex, Long-Running Tasks to Improve Agent Reliability and Focus #5996
2. Feature Request: Advanced Agent Orchestration with Hierarchical Delegation and Persistent Agents #5898

This issue will be automatically closed as a duplicate in 3 days.

• If your issue is a duplicate, please close it and 👍 the existing issue instead
• To prevent auto-closure, add a comment or 👎 this comment

🤖 Generated with Claude Code

[ANogin]

Found 2 possible duplicate issues:

1. Feature Request: A Structured Workflow for Complex, Long-Running Tasks to Improve Agent Reliability and Focus #5996
2. Feature Request: Advanced Agent Orchestration with Hierarchical Delegation and Persistent Agents #5898

Definitely very related, but not necessarily the same thing.

[coygeek]

Hey, this is a great point. I've definitely run into situations where a high-level plan is solid, but each step is its own mini-project that needs its own planning and execution loop. Decomposing complex tasks like that is key.

It looks like the Claude Code team has a feature that's very close to what you're describing, called subagents. I've been using them for this exact kind of hierarchical workflow.

From the docs, subagents are basically specialized AI assistants you can delegate tasks to. Each one gets its own context window, so it doesn't pollute your main conversation, and you can give them very specific instructions and tool permissions.

Here’s how I’ve been tackling this, based on what's in the documentation:

1. Create the High-Level Plan: Just like you said, I start by asking Claude to create the main plan, and it uses the TodoWrite tool to lay it out.

   > I need to refactor our authentication system to use OAuth2. Create a detailed migration plan.
   

   Claude might come back with a plan like:

   • 1. Analyze current auth implementation.
   • 2. Implement OAuth2 client logic in a new module.
   • 3. Create a feature flag to switch between old and new auth.
   • 4. Update frontend components to use the new flow.
   • 5. Add integration tests for the OAuth2 flow.

2. Delegate a Complex Step to a Subagent: Step 2, "Implement OAuth2 client logic," is definitely not a one-shot task. This is where I'd explicitly invoke a subagent. I could even create a custom one for this.

   > Okay, let's tackle step 2. Use the `api-designer` subagent to implement the OAuth2 client logic. Think through the edge cases like token refresh and error handling.
   

3. Hierarchical Execution: This is where the magic happens. The api-designer subagent would then kick in and create its own plan and todo list for just that task, without me needing to manage it. It would do its own file reads, writes, and tests in its isolated context. Once it's done, it reports the result back to the main agent.

4. Continue the Main Plan: Once the subagent is finished and I'm happy with the result, the main agent marks step 2 as complete and moves on to step 3 of the high-level plan.

The really powerful part is that you can create your own custom subagents for your project's specific needs. You can use the /agents command to create a new one, define its purpose, and limit what tools it can use. For example, you could have a database-migrator subagent that's only allowed to interact with schema files and the Bash tool for running migration scripts.

It seems to map perfectly to the "spawn a copy of itself as an agent" idea you had.

Check out the official docs on it, they have some good examples:

• Subagents Documentation: https://docs.anthropic.com/en/docs/claude-code/sub-agents
• Common Workflows (has a section on subagents): https://docs.anthropic.com/en/docs/claude-code/common-workflows

Hope this helps! It seems like a perfect fit for the problem you're describing.

[github-actions]

This issue has been inactive for 30 days. If the issue is still occurring, please comment to let us know. Otherwise, this issue will be automatically closed in 30 days for housekeeping purposes.

[ANogin]

I believe this is still relevant.