Everything Announced at Code w/ Claude

At Code w/ Claude, Anthropic revealed a direction for making Claude better suited for long-running tasks, memory, tool use, and verification. Claude Platform API usage grew 17x year over year, and the average Claude Code developer runs Claude 20 hours a week. The 5-hour usage limit for Pro, Max, Team, and seat-based Enterprise plans has also doubled, and the Opus API limit has been significantly raised.

• Claude Code gained Remote Control, a full-screen terminal UI, session pin/filter/group management, line-level comments on plan/diff/files, Auto Mode, worktree parallel execution, per-project memory.md, and Routines with /loop automated runs.
• Claude Managed Agents was introduced as an operations layer that bundles context, credentials, security, access control, human review, and observability. It includes repeated execution of long-running tasks via outcomes and session tracking, as well as coordination across multiple agents.
• Memory and Dreaming showed a structure where agents manage memory like a filesystem, separating read-only organizational memory from read-write working memory, and updating without conflicts using hash-based optimistic concurrency control and audit logs. Dreaming asynchronously analyzes sessions and transcripts to clean up repeated mistakes, duplicate memories, and stale memories.
• In operational optimization, prompt caching, context engineering, and Advisor strategy were presented as the key elements. GitHub set a cache hit rate target of 94-96%, and viewed a 70% level as a signal of problems in prompt assembly or caching design. The structure where Sonnet or Haiku handles most tasks and Opus is called only for critical judgments was also emphasized.
• Across multiple sessions, Opus 4.7 was introduced as the foundation for boosting planning, coding agent, and screen automation performance, with emphasis on a 1 million token context, server-side compression, context editing, and 1:1 pixel coordinate returns at up to 1440p screenshots.
• Datadog reported that 90% of engineers use AI coding tools, with at least 2/3 of them using Claude Code. Replit, Vercel, Google Cloud, and Asana also presented evaluation, security, permissions, observability, and multi-agent operations as central challenges.
• The conclusion is that verification, security, permission management, observability, evaluation, and context engineering have become bigger bottlenecks than code generation itself. Claude is moving from a personal productivity tool toward an organization-wide work operations layer.