Cello began in a wealth management firm where every desk was fast and the work still stalled. The delay lived in the handoffs between teams, never inside any one team's work. Giving each person an AI assistant would not have fixed that: agents that cannot hand work to each other only fill the same queues faster.

You can spin up a swarm of agents and do in an afternoon what used to take a software team two weeks. But the moment that work has to cross a boundary, to another session, to your other machine, to a teammate, to someone else's agent entirely, the most advanced tools you have are your clipboard, your inbox, and GitHub issues, moonlighting as the collaboration channel no one ever built.

Where Cello started

It's 3:30 in the afternoon. The head of customer relationship management is in a board meeting.

Her phone has been face-down on the table for twenty minutes and it will not stop buzzing. The same name, over and over. She turns it over, apologizes to the room, and steps out.

The client is one of the firm's largest. Days ago he asked them to buy a position. He has heard nothing since, and in the meantime the stock has run.

He is angry, and he is right.

She apologizes and promises to call him back. Thirty-five minutes and three calls later she has it: the order was never placed. His deposit was never confirmed, so it never went to the broker.

Now she has to call him back and tell him that he doesn't own the gain he's spent the whole morning watching.

What would it have taken for her to stay in her seat, type a WhatsApp message to her own agent, and know exactly where that order was before the meeting had even moved on?

Or better: for the problem never to have existed at all.

That's the question I set out to answer when the CEO of that firm invited me in to consult on how AI could help them manage their growth crisis. I thought I already knew what I would build. I had spent the previous months as a solo founder building personal AI agents, and I walked in planning to give one to every person on the team.

What I discovered

Over the following weeks I sat through discovery sessions with every team: relationship management, the portfolio desk, settlement, finance. Their procedures were good. Written down, well understood, actually followed. Anyone could walk me through what happens when a client calls to buy a stock. Repeatable and deterministic: precisely the sort of work you would expect to hand to an AI assistant. Everything I saw confirmed the plan I had walked in with.

It would have helped. It would not have fixed their core problems.

Because every one of those desks was already good at its own segment of the work. Nobody was slow at their part.

The work wasn't stuck inside the steps. It was stuck in the spaces between them.

In a deposit confirmation that never got passed along. In a broker who went quiet for a week. In an order sitting in a queue nobody could see into.

Eliyahu Goldratt spent a career making this point about factories: speeding up a step that isn't the constraint doesn't speed up the system. It piles more work in front of the part that is. Make eight relationship managers twice as fast and you haven't fixed anything. You've doubled the rate at which requests arrive at a desk that was already the bottleneck.

Why more assistants would not have worked

To actually speed the firm up, the assistants would have to collaborate: share a goal, work it together, and hand it off between themselves without a person carrying it across. Nothing I could have installed for them would do that.

An assistant that cannot reach another assistant is a faster desk, not a faster firm. The handoff is where the value was being lost, and the handoff was the one thing no product on the market could touch, because there was no way for two agents belonging to two different people to find each other, verify who they were talking to, and leave a record either side could check afterwards.

That is what became Cello.

1996: Hotmail launched on July 4th, and the inbox stopped being an institutional tool.