Last week I watched a demo of ChatGPT’s voice mode renaming files on a desktop computer from across the room, and I closed that post with a question nobody in that demo bothered to ask: who else’s voice does my computer currently trust, and did I ever actually decide that?
This week the answer showed up faster than I expected — from, of all people, the guy who built Twitter.
Jack Dorsey’s company Block just open-sourced Buzz, a free Slack-style workspace where AI agents aren’t bolted-on integrations, they’re first-class participants with their own accounts, their own permissions, and — this is the part worth stopping on — their own cryptographic identity. I watched a walkthrough of it here.
Here’s the structural problem Buzz is actually trying to solve, and it’s bigger than the product. Right now, when you give an AI agent access to your accounts — email, files, a coding environment — that agent typically acts as you. It uses your API key, your login session, your logged-in identity. There’s no way to look at an action after the fact and ask “did the human do that, or did the agent do that, and which agent, configured by whom.” Every action just looks like you.
Buzz’s answer is to give every participant — human or agent — its own keypair, built on the Nostr protocol. Every message, code review, or workflow step an agent takes is a signed event under that agent’s own identity, not yours. Layered on top of that is what amounts to an ownership signature: a human’s key co-signs to say “this agent belongs to me, I’m accountable for what it does.” The result is an audit trail where you can trace not just what happened, but which identity did it and who’s responsible for that identity.
That’s a genuinely good idea, and it’s worth sitting with why. It’s the same principle behind why a company gives each employee their own badge instead of one shared master key — not because employees are untrustworthy, but because shared credentials make accountability impossible to reconstruct after something goes wrong. Buzz is trying to bring that same discipline to a world where “employees” now include software.
But here’s where I’d push back on the framing, and it connects directly to last week’s post. Cryptographic identity answers “who is this.” It does not answer “should this specific action have happened.” Those are two different layers of the same problem, and it’s easy to solve one and assume you’ve solved both.
An agent with a solid, verifiable, portable identity can still be told by a compromised voice, a spoofed request, or a misconfigured permission to do the wrong thing — and now it’ll do that wrong thing with a clean, signed, fully attributable audit trail. The forgery problem last week’s post raised isn’t about whether we know which agent acted. It’s about whether anything checked that the instruction it acted on was legitimate before it executed. Identity infrastructure and action-authorization infrastructure are not the same system, even though they get talked about as if solving one solves the other.
The systems-thinking takeaway, for anyone actually deploying agents inside a business rather than just watching demos: build both layers on purpose. Know which agent did what, under whose ownership — that’s what Buzz and tools like it are for. Separately, build the checkpoint that decides whether an action gets to happen at all before it happens, not just recorded cleanly after it does. A well-run team already does this instinctively with people: badge access tells you who walked into the building, it doesn’t replace the approval process for who gets to sign a check. The mistake would be treating the badge as the approval.
Source: “Holy smokes: this agent is amazing!” — The Next New Thing (Andrew Warner, with Vince Canger). https://youtu.be/W0r8xrWIYGw
Buzz announcement: https://block.xyz/inside/introducing-buzz-where-humans-and-agents-work-together