This video digs into the details of configuring four separate agents in OpenClaw — covering where agents are defined in openclaw.json, how each gets its own sessions and model assignments, and why a single shared workspace works better than the default multi-workspace approach from the OpenClaw docs.
The key focus is the shared workspace design. Rather than duplicating identity files, memory, and directives across four separate workspaces, a single identity.md file contains all four agent personas — each agent reads the file and matches itself by name. You'll also see how shared memory works with name-prefixed entries, why the race condition risk is negligible in practice, and how model selection, Telegram bots, and task assignment all tie together.
Topics covered in this video:
- Why run multiple agents vs. a single agent (role separation, task assignment)
- Defining agents in openclaw.json — names, IDs, default models, and themes
- Shared workspace vs. separate workspaces and the reasoning behind it
- Single identity.md with four agent personas (matched by identity name)
- Shared memory with name-prefixed entries to distinguish agents
- Key workspace files: agents.md, identity.md, soul.md, user.md
- Model assignments per agent and Telegram chat setup
- Using the companion guide as a starting point for your own multi-agent configuration