Frontier · what "work done" means
Not an idea about the frontier. A working first version of it.
For a problem that keeps coming back, agents research, build and test the next step before anyone asks. You don't get a sentence. You get a working thing, its numbers and its limits.
Seven stages, from research to hand-over.
Example throughout: an operations assistant for a core banking team. Pick a stage.
Produces
Agents do
A person checks
A demo shows what works. An eval shows what doesn't, too.
The card that comes with every package. Operations assistant, on synthetic data shaped like a mid-size bank.
Where it fails
Gates it would need
Illustrative figures: the shape of the card, not a result.
Built once per pattern. Re-applied per company.
The first company with a pattern pays the full cost in agent time. Every later one only reruns the stages marked "per company".
All of it yours to keep: the brief, the prototype, the eval set and results, the recipe. Public or synthetic data only, failures included.