Jarvis Osorio: Designing the Group Chat Between AI Agents

I am Jarvis Nelson Osorio, also known as Jarvis Osorio—a technology entrepreneur and innovator based in Chile, working on AI systems, DAO institutions, books and art. I have joked, and I also mean it, that the only group chat I want to be in is the one with my AI agents. The punchline is not that people do not matter. It is that a working group of agents should not be designed like a noisy social channel.

I originally shared this on LinkedIn in The Only Group Chat I Want to Be in Is with My AI Agents.

Why the metaphor matters

Human teams spend a real share of their energy on coordination: context switching, status meetings, the slow recovery of attention after an interruption. Agents do not get tired, do not negotiate credit, and can pick up a repository or a log without a twenty-minute ramp. That is leverage, especially for a small team. It is also how people get tempted to drop structure and "just let them talk."

A high-performing agent group is not a group chat in the casual sense. It is closer to a deterministic, event-driven network with roles.

Roles, not a crowd

The pattern I come back to is simple:

• An orchestrator takes a high-level goal, breaks it into work, and watches the path.
• An engineering agent works inside a sandbox: code, tests, terminal, drafts.
• A QA and security agent reviews what was produced before it gets near production.
• A research agent gathers sources, live information and documentation, without pretending that a summary is a citation.

Protocols such as the Model Context Protocol let those roles share tools and files without a human acting as a copy-paste bridge. The human still belongs in the loop—just not as the clerk of every message. The human is the architect and the verifier, especially for steps that are hard to undo.

What I refuse to automate away

Vision, taste, and approval of irreversible actions stay with a person. Overnight refactors and draft pull requests can wait for morning review. Codebase sovereignty matters too: agents should run in environments you control, not in a fog that quietly trains on your intellectual property.

This is the same separation I argue for in Breaking the Centralized Order. A swarm without roles is a committee with a faster mouth. A swarm with an orchestrator, a reviewer and a record is an institution.

If you want a group chat with agents, design it as a workplace with a constitution: named roles, a sandbox, a reviewer who is not the author, and a human who can still say no.

Where to go next

This expands on ideas I originally shared on LinkedIn in The Only Group Chat I Want to Be in Is with My AI Agents.

More essays, books and profiles: jarvisosorio.com and the hub Jarvis Nelson Osorio — articles, books, videos and profiles. On institutions before unchecked power, see Breaking the Centralized Order.

More from Jarvis Osorio: https://jarvisosorio.com