MCP is the boring center that makes any of this work. The cliché is that MCP is “USB-C for AI”: one open protocol, any instrument. Like most analogies, it’s about eighty % proper. The half that issues is the eighty: I don’t have to jot down a customized adapter for each system the agent talks to. One MCP server per instrument and each agent talks to all of them the identical approach.
Typed handoffs between brokers are my very own structure, layered on prime of MCP somewhat than supplied by it. Every agent writes a typed artifact the following agent reads. Every handoff is logged with provenance. When one thing went incorrect six phases in, I may replay the chain. With out that self-discipline, a multi-agent pipeline is a debugger’s worst day. You realize the check plan is incorrect. You can’t inform whether or not the error got here from the Figma learn, the necessities interpretation or the ticket scaffolding. With it, I may level at precisely which stage went sideways and which inputs it was taking a look at when it did. The sample lives in a public MIT-licensed reference implementation for any reader who desires to run it.
The sixteen-minute quantity is the advertising and marketing quantity. I ran the complete chain finish to finish in about sixteen minutes on an artificial net-new display, Figma in, automation suite out. That repeated throughout my runs; it isn’t a demo trick. However sixteen minutes is the a part of the story most enjoyable to inform and least helpful to study from. It’s what will get quoted within the all-hands. The hours that come after, when a human opinions every handoff, are the place the work truly lives.
