Anthropic adds agent orchestration, Dreaming, and higher rate limits

Anthropic used its Code w/ Claude event on 6 May 2026 to focus on product updates around existing models, not on a new model release. In the morning keynote, Chief Product Officer Ami Vora said the goal was to make Claude’s products work better for developers, while also noting that API volume on Anthropic’s platform is up 17 times year on year.

The biggest near-term change for users is capacity. Anthropic said it is increasing rate limits for developers on Claude Code and the API, including doubling the five-hour limit for Pro, Max and Enterprise customers. The company also said it is partnering with SpaceX to use capacity from the Colossus data center, which it described as part of the infrastructure behind the expanded limits.

Anthropic’s product team also spent much of the keynote on how users can get more from the current models. Dianne Na Penn, head of product for research, said Claude’s abilities around tool use, long context, computer use, adaptive thinking, visual design and agentic loops are now strong enough to support more advanced workflows. She pointed to examples such as amp shifting its planning mode to Opus 4.7 and said teams getting the most value are using automated evals, simple scaffolding and novel model applications.

Ami Vora and other speakers framed one major theme as model routing, where a larger model such as Opus provides advice to a smaller one. Anthropic said this “advisor strategy” produced better benchmark results for Sonnet when it called Opus as an advisor, while also lowering cost. The company cited a customer named eve, which it said got frontier-model quality at five times lower cost.

The centerpiece of the platform update was Claude Managed Agents. Anthropic said the system is meant to help teams ship faster by bundling common agent patterns and controls, including memory. The company announced three new features: multi-agent orchestration, outcomes, and Dreaming. Multi-agent orchestration is meant for fleets of agents working on complex tasks, while outcomes lets users define what success looks like so Claude can iterate toward it.

Dreaming is the most experimental of the three. Anthropic described it as a research preview that can inspect previous sessions, identify what was missed, and create new memories overnight. In one demo, the system produced a descent-playbook.md file after reviewing prior work. Anthropic said multi-agent orchestration and outcomes are in public beta, while Dreaming requires access.

Later in the keynote, Cat Wu, head of product for Claude Code, outlined how Anthropic is expanding the product surface beyond the CLI. Claude Code began as a command-line tool, then added an IDE experience, and now has a desktop app for users who want a full-screen GUI with previews, images and rich outputs. Wu said both the IDE and desktop app are built on the same Claude Agent SDK that external developers can use.

Wu also announced Code Review, which Anthropic said is used by every team inside the company, along with Remote Agents, which lets users control their laptop from their phone. The company mentioned CI auto-fix, which can automatically patch failing pull requests, and Claude Security Reviews. Anthropic also said customers including Shopify and Mercado Libre are aiming for “90% autonomous coding by Q3 this year.”

Boris Cherny, who created Claude Code, closed the keynote with a demo of Claude working on a refund feature for an ACME dashboard. The demo included idempotency, multi-currency handling and audit logging, and showed Claude using the web UI directly to find an edge-case bug. Cherny said a lot of future code will be written asynchronously, with routines acting as higher-order prompts that can set up automations and return with PRs ready to merge.

Anthropic ended the keynote with a clear message: the company is not trying to sell a new model today, but to make the current stack more useful for software teams. That includes more capacity, more agent controls, more review automation and more ways to run Claude work in the background.