July 22, 2026
July 22, 2026
They called it the next unsolved problem. It had been running for months.
Two talks, a few months apart, describing systems that were already running here. What that is worth, and what it is not.
Two talks, a few months apart, describing systems that were already running here. What that is worth, and what it is not.
This is an awkward thing to write without sounding pleased with yourself, so here is the plain version first. Twice in the past year, a stage we were not standing on described a system already running here.
The first was Y Combinator, on building a company with AI from the ground up. The argument was that AI should not be a tool a company uses. It should be the thing the company runs on.
“Every important action should produce an artifact that the intelligence at the center of the company can learn from and use to self-improve.”
That is close to word for word the mission statement of the records system Siege runs on, written before the talk went up. Every action the company takes becomes a durable record that can be searched later.
The second one was more specific
Peter Steinberger, on an OpenAI panel this summer, described giving up on running ten terminal windows at once.
“I thought I was orchestrating. Really, I was polling. I was the scheduler, the router, and the memory.”
What he wanted instead was one long running manager that spawns workers, wakes on triggers, and can be steered from anywhere. Something you can text, rather than a session trapped inside an app.
He closed by naming the gap.
“Models are advancing faster than the harnesses and organizations around them. Designing those things is the next engineering problem.”
I had been running one for months. It dispatches work to a fleet of agents, checks the result at a gate before anything ships, and messages me when it lands. I steer it from my phone.
What that is actually worth
On its own, not much. Being early to an idea is a coincidence unless something came of it.
Here is what it is worth. Every company selling this kind of work has opinions about how it ought to be done. Very few of them are running the thing they describe, on their own operations, where a failure costs them instead of a client.
We are, and that can be checked. Almost nothing else in this field can.
Anyone can produce opinions indefinitely. Nobody can produce footage of a real system breaking and getting repaired unless they have one.
This is an awkward thing to write without sounding pleased with yourself, so here is the plain version first. Twice in the past year, a stage we were not standing on described a system already running here.
The first was Y Combinator, on building a company with AI from the ground up. The argument was that AI should not be a tool a company uses. It should be the thing the company runs on.
“Every important action should produce an artifact that the intelligence at the center of the company can learn from and use to self-improve.”
That is close to word for word the mission statement of the records system Siege runs on, written before the talk went up. Every action the company takes becomes a durable record that can be searched later.
The second one was more specific
Peter Steinberger, on an OpenAI panel this summer, described giving up on running ten terminal windows at once.
“I thought I was orchestrating. Really, I was polling. I was the scheduler, the router, and the memory.”
What he wanted instead was one long running manager that spawns workers, wakes on triggers, and can be steered from anywhere. Something you can text, rather than a session trapped inside an app.
He closed by naming the gap.
“Models are advancing faster than the harnesses and organizations around them. Designing those things is the next engineering problem.”
I had been running one for months. It dispatches work to a fleet of agents, checks the result at a gate before anything ships, and messages me when it lands. I steer it from my phone.
What that is actually worth
On its own, not much. Being early to an idea is a coincidence unless something came of it.
Here is what it is worth. Every company selling this kind of work has opinions about how it ought to be done. Very few of them are running the thing they describe, on their own operations, where a failure costs them instead of a client.
We are, and that can be checked. Almost nothing else in this field can.
Anyone can produce opinions indefinitely. Nobody can produce footage of a real system breaking and getting repaired unless they have one.
