Two agents, one project, no collision
A real overnight decision page, replayed: the night my parallel sessions started coordinating on their own.
Built 2026-09-28 · Filmed 2026-09-28
Just after midnight I asked a tab to keep working on a project while I slept. It noticed another tab already on the same project and split the work with it before asking me anything.
What you're watching
The clip opens on that page, and its warning box is the moment. The other tab had claimed one stage of the project. So this tab took two other stages, told the other tab in a single message and asked me to approve the plan before bed.
Then it steps back to my workflow map to show what made that possible. One command replaces the note I used to type into every tab before going to sleep.
It hands the project to the drive, the loop that runs a project stage by stage. Before every question it brings me, the drive checks which other tabs are live on that project and what each one has claimed. Any tab can also ask the overseer, a lookup over all live tabs, whether someone else touches what it is about to do.
Back on the page, I leave a comment anchored to the warning box, which goes back to the agent as work. Then I take the recommended answers and fill in the note I actually left that night. The voice drops out while the text is being written, so you can read it.
Spotting a collision is the easy part. The hard part is what came next: without being asked, one tab worked out a split with the other, and both kept to it while writing to the same project files. None of this is one feature. It is several small pieces composing. Sessions know about each other. A shared project plan lets each one claim a stage. A one-message rule replaces polling. And a single decision page appears only when a human is actually needed.
What doesn't come across on video: the page is a faithful replay. It is served from the saved copy of the original gate, with the same text and numbers. It runs inside a sandbox, so the replayed comment and answer land nowhere real. The note is mine, word for word. The comment is the one thing added for the demo, and it asks for something real: messages between tabs like this one are not yet logged anywhere the overseer can show them later.
Transcript
English
Just after midnight, I asked a tab to work through the night. It wrote me this page first. Before asking me anything, it noticed another tab was already working on the same project. That tab held stage four. So this one took stages 5b and 2b, told the other tab in one message, and left stage four alone.
How did it know? Overnight is one command I type before bed, in place of the message I used to write into every tab. It hands the project to drive project, which runs it stage by stage. Before each question for me, it checks which tabs are live here and what each has claimed. On a clash, it tells the other tab, once. And any tab can consult the overseer. Which live tabs touch this? Is someone already building it? Will my change collide?
It answers from a digest of every live tab, at zero model tokens. Back on the page, I can comment on any part of it. The comment stays anchored to that warning, and goes back to the agent as work. This one asks for tab-to-tab messages to be logged too.
The work itself is a UX tester I'm building. Tonight it tries a first-time-user persona on public sites, with a four-dollar cap. One click takes both recommendations. Then, the note I actually left that night. Both tabs kept working on one project. They worked out the split themselves; I just approved it. Even this video was made by the same system.
Português
Depois da meia-noite, pedi pra uma aba trabalhar a noite toda. Ela escreveu esta página antes. Antes de me perguntar qualquer coisa, ela percebeu que outra aba já trabalhava no mesmo projeto. Essa aba estava com a etapa quatro. Esta pegou a cinco B e a dois B, avisou a outra numa mensagem só e deixou a quatro quieta.
Como ela soube? O overnight é um comando que digito antes de dormir, no lugar da mensagem que eu escrevia em cada aba. Ele passa o projeto pro drive project, que roda etapa por etapa. Antes de cada pergunta pra mim, vê quais abas estão ativas e o que cada uma reservou. Se houver choque, avisa a outra, uma vez só. E qualquer aba pode consultar o overseer. Quais abas ativas mexem nisso? Alguém já está construindo? Minha mudança vai colidir?
Ele responde com um resumo de todas as abas ativas, a zero token de modelo. De volta à página, posso comentar qualquer parte. O comentário fica preso àquele aviso e volta pro agente como trabalho. Este pede que as mensagens entre abas também sejam registradas.
O trabalho em si é um testador de UX que estou montando. Esta noite, uma persona de primeiro acesso testa sites públicos, com teto de quatro dólares. Um clique aceita as duas recomendações. Depois, a nota que eu de fato deixei naquela noite. As duas abas seguiram no mesmo projeto. Elas mesmas dividiram o trabalho, eu só aprovei. Até este vídeo foi feito pelo mesmo sistema.