Features designed to guide earlier, less capable AI models, like "plan mode," are becoming obsolete. As models improve, they can execute complex tasks directly, making scaffolding features unnecessary and highlighting the rapid pace of model-driven product deprecation.
As AI models become increasingly powerful, the primary bottleneck shifts from the agent's capabilities to the user's ability to understand and leverage them. Building a user's mental model of what's possible can unlock more value than incremental UX improvements.
Startups that built complex, bespoke systems like RAG for early models now face a dilemma. These systems, once an advantage, can become technical debt as platform-native capabilities (like file search or simple `grep`) surpass them, making it difficult to stay on the frontier.
Formal evaluations ("evals") are not effective for zero-to-one innovation. Anthropic's Thariq Shihipar advises early-stage startups to avoid building complex eval systems, which slow them down, and instead iterate fast to build intuition about what works. Evals are for scaling, not discovery.
Designing the prompts, skills, and structures ("harnesses") that guide AI agents is not a standardized engineering discipline but an art form. It relies heavily on human taste and intuition, and there is no single consensus on the best approach, even among experts at a frontier lab.
Instead of providing a detailed but potentially incomplete prompt, ask the AI model to interview you about your goal. This meta-prompting technique forces clarification and helps uncover "unknown unknowns" you hadn't considered, leading to a much better final output.
A core principle at Anthropic is to challenge standard business trade-offs like "good, fast, cheap—pick two." Instead of accepting these as constraints, the team is pushed to find ways to achieve all desired outcomes simultaneously, forcing unintuitive breakthroughs rather than simple prioritization.
Even if model development stopped today, it would take up to a year to fully discover and utilize the existing latent capabilities. This "capability overhang" means the gap between what an AI can do and what we know how to make it do is massive, representing a huge opportunity.
Simply making your team more productive with AI (e.g., doubling PRs) won't increase revenue unless you redesign your business model to leverage that new capacity. The goal isn't to do old things faster, but to find entirely new things that are now possible, like letting customers order cars via email in 1995.