Anthropic and SpaceX: Strategic Alliance to Boost Claude Code

The "Code with Claude 2026" event by Anthropic was a pivotal moment for the company, featuring strategic announcements and captivating technological demonstrations. Right from the start of the day, at 08:56, participants took their seats in the main hall, ready for the keynote to begin at 9 AM. The opening was marked by a charming animation featuring the pixelated orange character Claude, which immediately captured the audience's attention.

At 09:05, Ami Vora, Product Director at Anthropic, took the stage. She replaced Mike Krieger earlier this year, who is now co-head of Anthropic Labs. Vora shared anecdotes about the rapid development of Scott MacVicar's teams at Stripe and Felicia Curcuru's teams at Binti, highlighting the efficiency of internal processes. However, she also expressed her hope to see announcements of new models or features, which did not occur during this session.

At 09:09, Vora illustrated the improvement of models by mentioning how Mythos discovered a 27-year-old vulnerability in the OpenBSD source tree. She also revealed that API volume has increased 17 times year-over-year on the Anthropic platform, an indicator of the growing adoption of their solutions. Although no new model was presented that day, Vora emphasized the importance of optimizing existing products for users.

At 09:11, updates were shared regarding Claude-managed agents, including multi-agent orchestration and Claude Code routines. Vora pointed out that most people will experience AI through the innovations developed on the Claude platform. At 09:12, she announced exciting news: an increase in rate limits for developers on Claude Code and the API, effectively doubling the five-hour limit for Pro, Max, and Enterprise customers. This increase is made possible through a partnership with SpaceX to utilize the capacity of the Colossus data center in Memphis, despite its controversial environmental record.

At 09:14, Dianne Na Penn, Product Manager for Research, took the stage to discuss innovations in multi-agent coordination and visual design. She highlighted the evolution of models, which can now operate for hours, compared to just a few minutes the previous year. Na Penn also spoke about the importance of using tools, long context, computer usage, adaptive thinking, visual design, and agentic loops. "The model's intelligence - the fundamental basis - has become strong enough to support all of this," she stated.

At 09:16, she discussed how amp changed its planning mode for Opus 4.7, emphasizing the importance of visual design in Claude Design. She mentioned that Opus 4.7 has a real flair for visual design, with context windows that seem endless when combined with high-quality memory. Multi-agent coordination allows for achieving large goals that could not be accomplished with a single instance.

At 09:19, she noted that around the same time last year, models could operate for just a few minutes. Today, many people have them running for hours. She also emphasized the importance of designing for the next model, building things that do not quite work today with the assumption that they will start functioning with a model upgrade in the future.

At 09:22, Dianne explained that teams making the most of Claude focus on automated assessments, simple scaffolding, and imaginative uses of models that others have yet to discover. At 09:23, Katelyn Lesse and Angela Kiang took the stage to talk about the Claude platform and how to achieve the right results from it.

At 09:25, they introduced "the advisory strategy," where Opus can provide on-demand advice to smaller models. They achieved better benchmark results for Sonnet calling Opus as an advisor, with both higher benchmarks and lower costs. One client, eve, achieved "frontier model quality at 5 times lower cost."

At 09:26, they discussed the difficulty of combining speed and scale, and how Claude Managed Agents is intended to help teams ship "10 times faster." It consolidates many best practices ready to use, such as memory.

At 09:28, they presented three new features for Claude Managed Agents: multi-agent orchestration, to create fleets of agents to solve complex tasks; outcomes, to define what success looks like so Claude can iterate and accomplish the task; and "Dream," where Claude can inspect its previous sessions and determine what it missed to improve.

At 09:30, they provided an example of building a hypothetical product to land drones on the moon, using multiple agents to accomplish this task, such as a Commander, a Detector, and a Navigator. They admitted that the demo was complex and hoped that detailed notes would be published after the session.

At 09:32, "Dream" was presented as truly interesting, allowing a task to run overnight that examines previous sessions and creates new memories, as in the example where it created a file descent-playbook.md. Multi-agent orchestration and outcomes are in public beta, while "Dream" is in research preview.

At 09:34, Cat Wu, Product Manager of Claude Code, took the stage to thank users for trusting Claude Code on their production databases when Sonnet 3.7 was their best model. She then presented the documentation on Dreams, explaining that access must be requested to try it, hence the term "research preview."

At 09:37, Wu explained that Claude Code started with the CLI, offering the most customization and control, then added the IDE, the same agents but in a user interface where code changes can be tracked more easily. The latest surface is Claude Code on Desktop, an interface for those who want a full-screen graphical interface with full-screen previews, images, and rich outputs.

At 09:37, she clarified that both the IDE and the Desktop application are based on the same Claude agent SDK that external developers can use themselves. She mentioned that users wanted to spend less time on code review, which led to the launch of Code Review, used by every team at Anthropic.

At 09:38, she introduced Remote Agents, allowing control of a laptop from a phone. She shared her personal experience using Claude Code for the web on her phone, eliminating the need to leave a laptop open somewhere.

At 09:39, she mentioned "self-correcting CI," which applies automatic corrections to PRs, although the only available documentation is an entry in the release notes. Claude's Security Reviews were also mentioned.

At 09:41, Wu praised some clients of Claude Code, such as Shopify and Mercado Libre, which has 23,000 engineers. These companies aim to achieve 90% autonomous coding by the third quarter of this year, an ambitious goal that underscores the confidence placed in Claude Code's capabilities.

At 09:42, she observed that executives and managers are getting hands-on with coding, as it doesn't take as much time to contribute meaningfully.

At 09:43, Boris Cherny, creator of Claude Code, took the stage, expressing his amazement at the current capabilities of the tool, even while working on it daily. He conducted a demonstration with the Claude desktop application, showing how Claude works to add refunds to the ACME dashboard, with idempotence to avoid double refunds, multi-currency management, and audit logging for the compliance team.

At 09:45, Cherny showed how he had multiple sessions running in the Claude desktop application simultaneously, allowing him to switch between them and see which ones require input. "We believe that in the future, much code will be written asynchronously," he stated.

At 09:46, he explained that much of his code is built by routines, which he described as higher-level prompts. "With Routines, developers can set up asynchronous automations and wake up with PRs ready to be merged," he explained.

At 09:48, he emphasized that the idea with self-correcting PRs is that "the person who owns the PR will never see a red X," with Claude inviting Claude Code to act autonomously.

At 09:49, the keynote session concluded, with the day's theme - not surprisingly for an event called "Code with Claude" - focused on learning the most effective ways to use existing models.

At 09:51, the agenda for the rest of the day was presented, and the live blog came to an end.