Pattern to steal

Inspired by the idea behind "Claude Managed Agents" style delegated agent workflows.

Problem

OpenClaw can already spawn subagents, ACP sessions, cron jobs, and durable task flows, but the user still has to think in low-level session mechanics more than they should.

There is a missing product layer for managed agents that feel like durable named workers with clear ownership, lifecycle, memory boundaries, and follow-up semantics.

Proposed feature

Add a managed-agents layer that supports:

• durable named agents / roles
• long-lived delegated work with explicit ownership
• clean task inbox / outbox semantics
• clear memory/context boundaries per managed agent
• visibility into status, blockers, last activity, and outputs
• easy escalation / handoff back to the user or another lane

Why this matters

This would make OpenClaw feel much more like an actual multi-agent operating system instead of a toolkit where the human has to keep recreating the manager layer manually.

Relationship to existing issues

• likely belongs under #2620 (orchestration and multi-agent operator UX)
• overlaps with chief-of-staff / orchestration UX, but is specifically about durable agent identities and lifecycle management

Acceptance ideas

• define the minimum managed-agent object model
• define how task assignment, lifecycle, and memory boundaries work
• define how managed agents differ from raw sessions/subagents
• produce a first implementation slice that is useful without requiring a giant new framework