Agents are good now. Give one a focused job that takes an hour and there's a decent chance it nails it. That part isn't the problem anymore.
The problem is everything bigger. Give an agent something that takes days, or weeks, and you watch it slowly lose the plot. It rereads things it already read. It retries things that already failed. It forgets why it made a decision two days ago. And eventually it tells you it's done when it isn't.
Most real work looks like that second kind. Building a product, running a marketing program, keeping a business moving. That gap is what I'm here to close.
#The mission, plainly
I want autonomous systems that can carry real work for days and weeks at a time, or run whole parts of a business, without falling apart, and without taking the wheel away from the person in charge. And I want to give the engines that do it away, in the open.
For that to be true, three things have to hold.
#It has to last
Long work shouldn't get harder just because it's long. That's the whole reason StateFrame is built on state instead of a chat transcript. The model gets the job as it stands right now, not a growing pile of everything that happened before, so turn 300 looks about the same as turn 3. In a 300-request run, every prompt fit inside the same fixed budget.
I'm not there yet, and I won't pretend otherwise. Today, really long tasks and very big changes can still stall out, and I say so on the StateFrame page. Weeks is the goal, not the claim. But the shape is right, and shape is the hard part.
#You have to be able to check it
I don't trust an agent that says it's done. I trust one whose work got checked.
So in StateFrame the model only proposes. The engine does the actual work, in a sandbox, on a copy of your code, and the model can't touch the tests that decide pass or fail. When it says it's finished, the engine runs your checks again and shows you the diff. Every request is saved, so if something goes sideways you can see exactly what the model was looking at when it happened.
#A person keeps the wheel
I'm a bit of a control freak. I suspect most people running a business are too, and honestly I think that's healthy.
Autonomy should take the busywork off your plate. It shouldn't take the decisions. That's how Momentum works: small, reversible stuff you've already set rules for, it handles. Bigger stuff comes to you. It only gets more freedom by earning it, it asks before taking it, and it loses it on its own the moment it starts getting things wrong.
#The bet I'm making
Here's what I actually believe, and I could be wrong about it.
Most of the gap between the small open models and the giant ones isn't in the model. It's in everything around it: what the model gets shown, how its work gets checked, and how a long task stays on track. When a model stumbles, my first question isn't "do I need a bigger model?" It's "what did the engine fail to give it?" The second question is "how small would I have to break this down for it to pass?"
That second one is where it gets fun. My theory is that for a lot of everyday Python work, a great tiny model, something like MiniCPM5-2B, can still get the job done if the work is cut into small enough steps. I haven't proven that yet. It's one of the next things I want to find out.
I've seen early signs that this works. In one set of tests, a single fix in the engine took a recurring failure at the finish line from 21 of 24 runs right to 24 of 24. Same model. Better frame.
And I'm a big believer in compounding, and not just with money. Going from 59% to 62% on something won't make headlines. Stack a hundred of those and the hard problems start getting easy. That's the plan: lots of small, measured wins, kept and built on.
If the bet pays off, people get to run serious, long-running agents on hardware they own, on models they choose, with every step on the record. That's worth spending the next few years on.
#Why give it away
Because this stuff should belong to the people using it. What I release stays open, and I put that in writing. That promise gets its own post.
If I'm wrong about any of this, you'll read about it here. Then you'll read about what I did next.