What was delivered
- A fresh Ubuntu server on AWS running OpenClaw, the open-source agent platform, in its February 2026 release.
- Several agents on one machine, each with its own character, role and instructions: Tiger the team leader, Owl the designer, Monkey the developer, and more.
- A different cloud model behind each character, such as ChatGPT, MiniMax and Grok, so each role uses the model that suits it.
- One Telegram group as the workroom: the human and all the agents see the same conversation and answer each other in it.
- A leader that supervises: Tiger splits the work, watches the discussion, pulls an agent back when it drifts, checks the result against the request and reports to the human.
How a job runs
- I write the command in the Telegram room, in plain words.
- Tiger breaks it into tasks and assigns them by role.
- The agents discuss in the open: the designer proposes, the developer raises what is hard to build, and they agree.
- The developer builds; Tiger steps in when someone goes beyond the brief.
- Tiger reviews the result against the original request, then hands it to me for approval.
Why it matters
One assistant answers questions. A team with roles, a leader and an open room behaves like a small project team: the work is split, ideas are challenged, and someone is responsible for the result. Because everything happens in a chat group, I can follow it, or step in, from my phone.
It was my first hands-on multi-agent setup, and the lessons from it, about roles, supervision and keeping a human in the room, shaped the agent work that followed.