Thread - a quieter inbox for AI agents.
Reframing agent output as letters, not chat, so the work feels considered rather than reactive. A six-month design study, now in private beta with twelve teams.
Model memory drift in onboarding
Three runs disagree on when to ask for intent. Suggested margin notes attached.
Weekly synthesis from support
Two repeated complaints, one likely copy problem, no incident.
Pricing page experiment readout
Signal is real but thin. Keep collecting until Friday.
Model memory drift in onboarding
The agent found a difference between what users say they want and what the current prompt stores. It recommends splitting first-session intent from long-term learning goals.
01 The brief
A founder I'd worked with before came back with a complaint that had the shape of a real problem: her team was drowning in agent output. They had built a clever orchestration layer, but the UI was still a chat thread, and chat threads are not good at carrying the weight of work.
The ask was vague on purpose: make it feel less frantic. I took that as license to question the chat metaphor entirely.
"We do not actually want to talk to it. We want it to leave us something we can come back to." - first interview, week one.
02 What I tried first, and why I threw it out
The obvious move is a kanban: agent output as cards in columns. I built it in a week. It was fine. It was also more software about software, and the team already had three tools that looked like that. I shelved it.
The shift was noticing that nobody opened the agent's output the moment it arrived. They batched. They opened things in the morning, in a quiet half-hour, with coffee. The metaphor was not a chat. It was a stack of mail.
03 The shape it took
Thread is, at its core, three small ideas glued together:
- Agents write letters, not messages: a subject line, a body, a signature, an ask.
- Letters arrive on a tray, sorted by how much patience they require, not by when they arrived.
- You reply by annotating in the margin, like a manuscript. The agent reads the annotations and tries again.
None of this is technically novel. The novelty is in what we removed: notifications, unread counts, the idea that anything is ever urgent.
04 What I learned
The thing I keep coming back to: chat is a default, not a decision. We reach for it because it is easy to build, not because it serves the work. Agents that produce real output deserve real surfaces: surfaces that let the output sit, breathe, and be considered.
I am continuing to work on Thread with a small group of teams through the rest of the year. If your team is making something agent-shaped and you want to talk about the surface around it, write to me.