People ask us where Estelle sits against Devin or Cursor or Claude Code, and the honest answer is that it doesn’t sit against them at all. We use them. Our own engineers run them every day. We are building the thing they stand on.
That sounds like positioning, so let me explain how we got there.
Watch where they actually fail
Spend a week reading agent output on a codebase with real history and the failures sort themselves into a pattern. The model writes against an interface that doesn’t exist. It re-solves something the team settled in March and wrote down in a thread nobody fed it. It ships a change that passes the test it wrote and breaks a caller it never opened.
Look at those closely. None of them is a reasoning failure. The model reasoned fine. It reasoned about a world it could not see, and when you cannot see something, you fill it in with whatever is most likely. That is not a bug in the model, it is what any system does under that much uncertainty.
A smarter model fills the gap more convincingly. It still fills the gap.
So we went down a layer
What your agents are missing is not intelligence. It is the boring institutional stuff: what the code actually contains right now, what your team already decided and why, what broke last time someone tried this, whether the change in front of you holds up against the rest of the system.
Nobody is building that. Every serious lab is pointed at the model. The ground the model walks on is assumed.
We are building the ground.
Why this is the better bet anyway
Here is the part that convinced us commercially. The best agent today will not be the best agent next quarter. Teams will switch, repeatedly, probably to a different lab each time. If we had bet on being the best agent, every release from a frontier lab would be an attack on our business.
Because we bet underneath, the opposite happens. Every improvement anyone ships makes us more useful. We want the race to continue. We want it to be brutal.
We do not need to win it.