When Castform set out to build onboarding for their RL post-training platform, they made a critical assumption that nearly sank them: developers would actually use the interfaces they built. After testing notebook templates and a no-code config wizard with live code previews, the team realized nobody was reading any of it. "Devs open their coding agent and ask it to do everything," the company explained in a detailed post-mortem. "They don't read code or docs, they don't reach for a clean UI, they point Claude at the platform and expect it to figure out the setup." The result: Castform gutted their entire onboarding flow and rebuilt it around one core principle—assume your user's coding agent is the interface.
Two Approaches That Bombed
Castform's challenge was designing for two radically different developer profiles. On one side, experienced ML researchers who just needed to validate the service worked. On the other, average AI engineers who'd never post-trained a model and had zero appetite to spend days figuring it out. Their first attempt—Python notebook templates popular in research circles—was immediately dead on arrival. "Newcomers couldn't understand enough from glancing at a wall of text and code how this would fit their use case," the team noted. They pivoted hard to a no-code config wizard with live Python previews that updated as users filled out forms, plus one-click export to actual SDK code. Sophisticated on paper. Dead in practice.
The Revelation That Changed Everything
Two uncomfortable truths emerged from their beta launch. First: every hour spent polishing dropdowns and form fields was wasted effort—users weren't touching the UI anyway. "Even two well-crafted sentences were 'too much text'," Castform observed. Second, and more damning: time-to-value stretched beyond 30 minutes when you factored in designing a training run for your specific task, launching it, and waiting for enough steps to generate convincing results. By that point, users had already bounced. The 'aha moment' needed to arrive in under ten minutes, not after half an hour of patience-testing setup screens.
One Command, Plain Language, Done
The solution Castform landed on is brutally simple: run one command, copy a short prompt, edit it in plain language to describe your task, and let your coding agent handle the rest. In five to ten minutes, that agent designs an environment for your task, generates training and eval examples, writes reward functions (or sensible defaults), runs rollouts on Castform's platform, and reports back what it built along with baseline scores. The user doesn't get a finished run—that takes hours and GPU cycles—but they walk away with working setup code they understand and enough results to think "this is real, this could solve my problem."
The Deeper Shift: Your AI Agent Is the UI Now
"The deeper shift is that the user's coding agent is the interface," Castform writes. "Our job moved from building screens to handing the agent the context it needs to do this well, then staying out of its way." This isn't just a UX optimization—it's a fundamental reframe for how developer tooling gets consumed in 2026. If your onboarding flow requires human eyes scanning documentation or clicking through form wizards, you're adding friction that agents won't tolerate and users increasingly won't either. The platform still provides full observability surfaces (reward curves, rollout inspectors, interactive playgrounds) via their example results—but those are discoverable by the agent, not forced on humans as prerequisite reading.