Descriptions:
At Anthropic’s Code with Claude conference, Lucas — a Research PM at Anthropic — delivers a structured overview of how Claude’s built-in capabilities have absorbed what developers previously had to engineer themselves. The organizing thesis: the scaffolding that required hundreds of lines of custom code in 2025 now ships natively with the model, letting developers focus on product outcomes rather than reliability infrastructure.
The talk covers four major capability areas through before-and-after comparisons. On tool use, Claude now searches through full tool sets autonomously with high enough accuracy that pre-filtering routers typically degrade performance — they’re no longer needed. Context management now includes automatic compaction that discards large tool results while preserving decisions and key analyzed outputs, reducing token consumption mid-session. A hosted code execution sandbox now handles the entire write-run-fix loop inside a single API turn, eliminating round trips to a developer-managed VM. Computer use has been simplified with on-model coordinate calculation removing the need for custom coordinate math layers.
Practical tips accompany each section: adding output schemas to tool descriptions to save agent round trips, configuring pre- and post-tool-use hooks in Claude settings, using /context in Claude Code to inspect a live colored breakdown of context window usage, and /schedule for setting cron-triggered background agents. This session is especially useful for developers actively building on the Claude API who want to understand which previously necessary infrastructure they can now safely remove.
📺 Source: Claude · Published May 08, 2026
🏷️ Format: Keynote Launch
