Anthropic's London event produced a striking data point: roughly half the room had shipped Claude written code without reading it first. The company's new 'dreaming' feature is its answer to what that future actually requires.
At Anthropic's Code with Claude event in London earlier this week, Jeremy Hadfield, an engineer at the company, asked the packed room a simple question from the main stage: who here has shipped a pull request in the last week that was written entirely by Claude? About half the hands in the room went up.
Then he asked the follow-up. Who here shipped that code without reading it first? Most of the hands stayed up. The room did not react with surprise.
That was the signal.
Boris Cherny, who leads Claude Code at Anthropic, described what he saw in that moment during his opening keynote. The default in software development, he said, is no longer a developer prompting an AI for help. The default, he said, is now having Claude prompt itself. The framing was not presented as a warning. Anthropic's stated goal is to push automation as far as it will go: not AI that generates code for human review, but AI that checks and corrects its own work, handles its own error messages, and ships when everything runs as it should. As Ravi Trivedi, an Anthropic engineer, put it in a separate talk: the key principle is getting out of Claude's way.
The behavioral data from the London event suggests that message is landing. The hand-raise was not a controlled survey — it was a show of hands in a self-selected audience of developer-tool enthusiasts — but the directionality is not ambiguous. When half the developers at a conference explicitly designed around AI coding tools admit they have shipped AI-written code unread, the relevant question is no longer whether this will become normal. It is what 'normal' actually means for code quality, security auditing, and the role of the human developer in the loop.
The infrastructure answer Anthropic is building toward that question is called dreaming.
Dreaming is a feature for Claude Managed Agents that Anthropic shipped on May 6 at its San Francisco Code with Claude conference. The mechanism: between active work sessions, a background process reviews what an agent did — what it wrote, what it learned, what went wrong — and consolidates patterns into persistent memory. It is not session-level context. It is cross-session, cross-task memory consolidation. It surfaces recurring mistakes, convergent workflows, and shared team preferences. It restructures memory so it stays high-signal as it evolves. Anthropic describes it as especially useful for long-running work and multi-agent orchestration, where multiple agents need to build on what earlier agents learned.
This is the infrastructure layer that shipping unread code makes necessary. When a human developer never reads the code that ships, the knowledge of why the code is shaped the way it is lives nowhere in the human's head. The agent that wrote it has to carry that knowledge forward. Dreaming is the first named, productized attempt by a major AI vendor to solve that problem at the platform level.
The feature is currently in research preview — available by request to developers, but not in general availability. That matters for calibration. This is a concrete architectural direction, not a shipped product that teams can deploy into production today. The capability exists; the availability does not.
Alongside dreaming, Anthropic shipped two other features at the same event that moved from research preview to wider release. Outcomes lets developers specify a target result and run an agent in a loop until it achieves it — a self-grading mechanism the agent applies to its own work. Multi-agent orchestration allows a coordinator agent to spin up parallel subagents. Both are available to developers building on the Claude Platform now. Anthropic also doubled five-hour usage limits for Pro and Max subscribers and raised API rate limits for some tiers by up to 17 times, responding to sustained demand that has strained its infrastructure.
The countervailing view is not absent from the conversation. Outside the conference walls, developers on forums including Reddit and Hacker News have reported that AI coding tools are being pushed by managers chasing productivity metrics, while the technology itself adds review overhead that offsets the speed gains. Researchers have warned that AI tools can produce code with security vulnerabilities that human review would catch. And there are reports, including from 404 Media, that some developers' coding abilities have declined as they hand more tasks to AI. Katelyn Lesse, Anthropic's engineering lead, addressed the concern directly in an interview with MIT Technology Review: the old software development best practices still apply, she said, but she acknowledged that some teams may have lost sight of them.
Angela Jiang, Anthropic's product lead for the Claude platform, was more direct about the destination. The end state the company is working toward, she said, is Claude essentially being able to build itself.
The dreaming feature is the most concrete articulation of what that actually requires. If the developer is not reading the code, the agent has to remember what the code means. That is not a metaphor. It is an infrastructure design constraint — and it is now on the roadmap as a named product.
Sources: Anthropic's Code with Claude showed off coding's future—whether you like it or not — Will Douglas Heaven, MIT Technology Review, May 21 2026. Anthropic's Claude Managed Agents can now dream, sort of — Samuel Axon, Ars Technica, May 6 2026. Inside Anthropic's 2026 Developer Conference — Dan Shipper et al., every.to, May 7 2026 (updated May 23 2026).