Anthropic Doubles Claude Code API Limits, Adds Self-Improving & Multi-Agent Tools

TL;DR

• Anthropic has doubled API limits for Claude Code and raised limits for Claude Opus, enhancing scalability for developers.
• New Claude Managed Agents capabilities—Dreaming, Multiagent orchestration, Outcomes, and Webhooks—were introduced.
• The Outcomes feature notably improved task success by up to 10 points on internal benchmarks for complex problems.
• Recordings from the recent Code w/ Claude SF 2026 conference are available, with future events planned for London and Tokyo.

Anthropic recently concluded its annual developer conference, Code w/ Claude SF 2026, held in San Francisco, bringing together engineers and founders to explore the evolving landscape of AI-native engineering. The event showcased how the gap between an initial idea and production-ready software is rapidly narrowing, emphasizing the importance of designing for the "AI exponential." Industry leaders and customers, including Asana, Cursor, GitHub, Replit, and Vercel, shared insights into their experiences building with Claude, illustrating real-world applications and agentic development strategies.

A major announcement at the conference was the significant increase in API access for developers. Anthropic doubled the rate limits on Claude Code and substantially raised API limits for Claude Opus. These changes are already live, enabling startups, developers, and enterprises to build and deploy their AI solutions with greater reliability and at a larger scale, facilitating more ambitious projects and seamless integration into existing workflows.

Further expanding developer capabilities, Anthropic unveiled new features for Claude Managed Agents on the Claude Platform. These cloud-hosted agents are designed for teams looking to build and deploy sophisticated AI at scale. Among the new introductions is Dreaming, a scheduled process that reviews past agent sessions, identifies patterns, and refines the agent's memory, allowing for continuous improvement between runs by incorporating recurring mistakes and shared preferences into a more effective memory store. For more detailed information, developers can refer to the Claude Platform Managed Agents overview.

Another powerful addition is Multiagent orchestration, which allows a primary agent to delegate tasks to specialized subagents. These subagents can work in parallel on a shared filesystem, each with its own model, prompt, and toolset. The entire collaborative workflow remains traceable within the Claude Console, providing transparency and control over complex operations. This feature opens doors for highly modular and efficient AI systems tackling multi-faceted problems.

The introduction of Outcomes significantly enhances agent reliability and performance. Developers can now define a rubric specifying what constitutes a successful output. A separate grader evaluates each result within its own context window, guiding the agent to revise its work until it meets the predefined standard. Internally, this feature has demonstrated remarkable success, boosting task success by up to 10 points on the most challenging problems. Complementing this, Webhooks allow agents to run autonomously, sending notifications once a defined outcome is achieved, streamlining automated workflows.

For those who missed the San Francisco event, keynotes and breakout sessions from Code w/ Claude SF 2026 are now available on YouTube. Anthropic is also taking the conference experience global, with upcoming Code w/ Claude events scheduled for London (May 18-19) and Tokyo (June 5-6). Attendees can expect live streams of Day 1 keynotes and breakout sessions, along with new technical tutorials and customer stories inspired by the talks, extending the learning and networking opportunities for the global developer community. Check out the San Francisco agenda for a taste of what to expect.

Summary

• Anthropic significantly boosted developer capacity by doubling API limits for Claude Code and increasing limits for Claude Opus.
• The new Claude Managed Agents offer advanced features like Dreaming for self-improvement and Multiagent orchestration for complex task delegation.
• The Outcomes feature dramatically improves agent reliability, increasing task success by up to 10 points on difficult benchmarks.
• Developers can explore these new tools and learn from conference sessions, with upcoming Code w/ Claude events in London and Tokyo.

Source: Code w/ Claude SF 2026: Building on the AI exponential | Claude