On site, a job isn't done until someone's checked it. I run my AI coding agents the same way.
One agent can edit. The reviewers are read-only. A coordinator hands out one task at a time.
Building apps this way, mainly a roster tool for now, and posting what the checks catch.
@isaac_dev0 tbh the moat might just be that nobody wants to rebuild the context management bit once they see how fiddly it is. Have you tried swapping models mid-project, and did it hold up?
My map builder, two days apart. Tuesday's trees were green clouds on sticks, and the cottage looked like it was melting off the hill. Now the trees have actual branches, the roof is proper thatch, and the path looks like somebody has actually walked on it. Building with Fable 5.2
I saw a dev whose agent deleted node_modules twice while 'tidying up'. That's why my agents have job descriptions: one edits, reviewers are read-only, and a coordinator hands out one task at a time. Nothing's done until it's checked. What job would you never let your AI do alone?
@Galo2027 Open-world crime game in Unreal Engine 5. Claude's building it on a US$300 budget, $55 in and the car already does 0-100 km/h in 7.4 s. The buildings are still placeholder boxes, but the car's ready to get pulled over.
Started my open-world crime game in Godot, decided it looked too basic and moved the lot to Unreal Engine 5. Gave Claude a US$300 budget. Movement, swimming, ragdoll, camera and driving are done for about $55. The buildings are still placeholders, so don't get excited.