Cat Wu and Thariq Shihipar describe how stronger Claude models changed software work from closely supervised implementation to higher-level delegation. Faster execution raises the value of product taste, business judgment, and ambitious problem selection. It also makes previously risky choices, such as codebase rewrites, more practical when a strong test suite preserves behavior.
Claude Tag extends coding agents into shared Slack channels. It can remember channel preferences, monitor recurring work, open pull requests, search internal context, and let engineers, designers, and product teammates steer the same session. Anthropic says its internal version lands a majority of product-engineering pull requests, while complex or sensitive changes still receive ownership-based human review.
Evaluation is the foundation for increasing autonomy. Anthropic combines external and internal capability tests with behavioral evals, incident-derived regression cases, code review checks, and red-team environments. New models must pass the established suite before replacing an older model, and automated review expands only in code areas where the evidence shows it catches the failures the team cares about.
The conversation also covers prompt and tool design. Frontier models need fewer rigid examples and prohibitions, so Claude Code now uses model-specific system prompts and gives stronger models more context and judgment. Tools are kept distinct and general where possible, while UI-specific tools remain when they provide useful permission and review surfaces.
For long-running autonomy, Auto Mode classifies requested actions using the conversation, user instructions, sandbox boundaries, and permission context. Anthropic combines that layer with credential injection, identity separation, model alignment, and adversarial testing. The speakers still frame security as layered risk reduction rather than perfect protection, and they see human value shifting toward direction, quality, and deciding what is worth building.
Watch on YouTube
