Feature request: Interactive step-into mode for Task subagents

Problem

When Claude Code spawns a Task (subagent), the user loses all control until it returns. You specify a task, wait, and hope the results match what you wanted. If they don't, you start over with a better prompt.

This fire-and-forget model fails for anything requiring nuance because users can't:

• Course-correct mid-exploration ("not that, focus on X instead")
• Guide the agent's investigation in real-time
• Control what ends up in the summary returned to the parent

The current design conflates context isolation with loss of control. These are orthogonal concerns:

• Context isolation is a technical constraint (tokens are finite, exploration pollutes focus)
• Control is a human need (we steer, we course-correct, we collaborate)

You can have both.

Proposed Solution

Add a "Step into" option when a Task is invoked:

[Approve]  [Reject]  [Step into]

• Approve = fire-and-forget (current behavior)
• Step into = enter the subagent conversation interactively

When stepping in:

• User enters an interactive session with the subagent
• Context remains isolated (separate thread/memory)
• User can guide the exploration, ask follow-ups, redirect
• /return exits and sends a summary back to the parent context

The context stays isolated. The human stays in control.

Proof of Concept

We implemented this for deepagents-cli (LangChain's agent framework):

PR: https://github.com/langchain-ai/deepagents/pull/829**

The implementation adds:

• "Step into" option in the HITL prompt for the task tool
• Context stack to track nested conversations
• /return, /summary, /context commands
• Summary file that gets injected to parent on return

~450 lines of changes. Works well in practice.

Why This Matters

Subagents are powerful for context isolation, but the fire-and-forget model makes them unreliable for complex tasks. Users end up avoiding subagents entirely or re-running them multiple times until they get lucky.

Interactive step-into transforms subagents from "batch jobs wearing agent costumes" into actual collaborative tools.

---

Context isolation is about memory. Control is about agency. They were never the same thing.