Building a multi-agent stack: how ten agents share one brain

It's easy to bolt ten prompts together and call it a “multi-agent system.” The result is incoherent: the social agent contradicts the content agent, the email doesn't match the brand voice, and nothing builds on anything else. A real stack has to feel like one department, not ten strangers.

Every agent in our stack reads from the same workspace knowledge — the brand voice profile, prior approved work, connected channels, and audience notes. So the social agent writes in the same voice the content agent used, because they're both grounded in the same source of truth rather than each inventing a tone from scratch.

Agents pass work to each other as structured outputs. An SEO audit becomes the input to a content refresh plan; the content plan becomes the brief for a social batch. Each handoff carries the context forward, so the last agent in the chain is still working from the first agent's findings — the way a good team briefs the next person instead of starting over.

Above the agents sits a Chief of Staff layer that reads the whole workspace and proposes the day's work: what's overdue, what's waiting on review, which routine is due. It never runs anything on its own — it lays out the plan and you click to act. That's the difference between a pile of tools and a system that tells you what to do next.