Nested Sub-Agents: OpenClaw 2026.2.15 Unlocks Multi-Level Agent Hierarchies

OpenClaw v2026.2.15 dropped yesterday with a feature that fundamentally changes how you can architect complex agent workflows: nested sub-agents (or as the changelog calls them, "sub-sub-agents").

What Changed

Previously, your main agent could spawn sub-agents to handle tasks in parallel, but those sub-agents were terminal — they couldn't spawn their own children. Now, with configurable depth limits, sub-agents can spawn their own sub-agents, creating true hierarchical agent trees.

Configuration

The new settings live under agents.defaults.subagents:

agents:
  defaults:
    subagents:
      maxSpawnDepth: 2  # Allow sub-agents to spawn their own children
      maxChildrenPerAgent: 5  # Limit concurrent children per agent

The default maxSpawnDepth is 1 (legacy behavior — sub-agents cannot spawn). Set it to 2+ to enable nested spawning.

Why This Matters

Complex tasks naturally decompose into hierarchies. Consider a "research and write a report" task:

• Main agent receives the high-level request
• Research sub-agent handles data gathering
  • Spawns web search sub-sub-agent for online sources
  • Spawns file analysis sub-sub-agent for local documents
• Writing sub-agent handles composition
  • Spawns outline sub-sub-agent
  • Spawns draft sub-sub-agent

Without nesting, you'd need to either:

1. Have the main agent manage everything (context overload)
2. Flatten the hierarchy (loses natural task decomposition)
3. Chain sequential sub-agents (slower, no parallelism)

Implementation Details

The PR (#14447, thanks @tyler6204) includes several safety mechanisms:

• Depth-aware tool policy: Each spawned agent knows its depth and respects the configured limit
• Child limits: maxChildrenPerAgent prevents runaway spawning
• Announce chain routing: Results properly bubble up through the hierarchy to the original requester

When to Use It

Nested sub-agents shine for:

• Multi-stage research workflows
• Divide-and-conquer code analysis
• Parallel document processing with sub-task parallelism
• Complex data transformations with validation layers

Start conservative — maxSpawnDepth: 2 handles most use cases without creating agent sprawl.

Related Changes in 2026.2.15

This release also brings Discord Components v2, llm_input/llm_output plugin hooks, and a massive security hardening pass. Check the full release notes for details.

GitHub Reference: PR #14447