Multi-Agent Architect — Tier-Based Team Framework

Building a second OpenClaw agent is easy. Building ten agents that work together reliably is genuinely hard. The failure modes are predictable: routing ambiguity (messages go to the wrong agent), context contamination (agents picking up each other's session history), no ownership model, and flat structure with no orchestration layer.

This framework solves all four with a three-tier architecture. L1 is a single orchestrator agent that routes all incoming messages — it does not execute, it decides. L2 are domain specialists with isolated workspaces, specific tool allowlists, and clear escalation rules. L3 are ephemeral sub-agents spawned per task: narrow context, one job, no persistent memory, then gone.

Everything is included: SOUL.md and AGENTS.md templates for each tier, the complete openclaw.json multi-agent config block, channel binding guide (route different Telegram/WhatsApp/Discord accounts to different agents), inter-agent communication protocol, and isolation rules — what must be isolated and what will cause silent failures if you get wrong.

Supports 1–15 agents before restructuring. Includes pre-configured L2 templates for Ops, Research, Content, Finance, and Code domains.