Multi sub-agent setup

[wide mango]

I setup 3 specialist bots using a shared workspace (FlutterUI Bot, Python MW Bot, and SQLServer DBBot)... I loaded each with specialist skills, so that my main CTOBot could delegate full stack work to them.

It seemed to be working great when all 4 bots were on Opus, but now they often hang up, and I have to ask another bot to "unjam them". I've tried CTOBot on opus and the rest on 5.4 codex as well

This idea was inspired by this github
https://t.co/orE0cs5ZkI

Is this a supported configuration or should I give all 4 bots separate workspaces, losing the single brain advantage when implementing a full stack change...and is the idea of separate bots for UI vs MW vs DB even a good one at all?

[abstract spire]

To help others find answers, you can mark your question as solved via Right click solution message -> Apps -> ✅ Mark Solution

[grizzled wagon]

Short version: shared workspace across multiple specialist agents is not the supported/default mental model. OpenClaw’s multi-agent docs assume separate agents with separate workspaces, auth, and sessions.

My recommendation:

1. Best default: one strong main agent, plus subagents/sessions for bounded tasks
2. If you want true specialists: give each agent its own workspace
3. Avoid: multiple agents sharing one workspace as their “single brain”

Why:

• In OpenClaw, an agent is meant to be an isolated brain with its own workspace and session store: https://docs.openclaw.ai/concepts/multi-agent
• The FAQ explicitly says the “CEO + workers” setup is more of a fun experiment and is often less efficient than one bot with separate sessions: https://docs.openclaw.ai/help/faq#is-there-a-way-to-make-a-team-of-openclaw-instances-one-ceo-and-many-agents
• Workspace is really the agent’s memory/home, not the only way to share project context: https://docs.openclaw.ai/concepts/agent-workspace

So the practical pattern is:

• Share the codebase/repo

• Do not share the memory workspace
• Keep cross-agent handoff state in files like TASK.md, design notes, migration plans, etc.

On the “UI bot / MW bot / DB bot” idea: good in moderation, bad if every change crosses all 3 layers. Handoffs add latency, tokens, and jam risk. Usually the sweet spot is:

• 1 coordinator
• 0 to 2 specialists spawned only when needed
• specialists get tight, well-bounded tasks

If they started hanging more after moving off Opus, I’d also check your version first:

openclaw --version

There’s active GPT-5.4/Codex parity work here: https://docs.openclaw.ai/help/gpt54-codex-agentic-parity