Significantly lowers the barrier and cost of building custom software by automating mid-level engineering tasks.
What did Anthropic actually demonstrate at the Code with Claude event?
Anthropic held a two-day developer event called Code with Claude in London on May 19 and 20, 2026, and what it showed was not a new product but a behavioral shift that is already locked in across the industry. Jeremy Hadfield, an engineer at Anthropic, opened the main stage by asking the room how many developers had shipped a pull request last week written entirely by Claude, and almost half raised their hands. When he followed up by asking how many shipped without reading the code at all, most of those hands stayed up. Spotify, Delivery Hero, Monday.com, and several agentic development startups all presented how they have restructured their workflows around Claude Code. This was not a launch event, it was a documentation event, and what Anthropic documented is that handing code production to an agent without oversight has already become standard behavior for a significant portion of working developers.
How does the dreaming feature change what Claude Code can do?
The most significant technical announcement connected to the event was a feature called dreaming, which Anthropic released approximately two weeks prior to London. Rather than starting each new coding session from scratch, Claude Code agents now write notes to themselves as they work, capturing context, decisions, and observations specific to each task. A separate consolidation system reads all those notes across different sessions, identifies patterns, and makes that consolidated knowledge available to future agents working on the same codebase. The practical result is that each subsequent agent arrives with context, learns from prior errors, and ramps up faster than it would on a cold start. Dreaming does not make Claude Code more capable at any single task, it makes the system progressively better at your specific codebase over time, which is a fundamentally different kind of improvement than a model upgrade.
Should small business owners care about the Code with Claude event?
Operators who pay for custom software or maintain a development team should care because the cost basis of mid-level coding work is being compressed by the same tools the industry just celebrated in London. Katelyn Lesse, Anthropic’s Claude engineering lead, stated at the event that Claude is currently “probably as good as a midlevel engineer at writing code” and confirmed that most of Anthropic’s own software, including Claude Code itself, is now written by Claude. For operators tracking how the cost of technical execution continues to shift, agentic coding signals and automation breakdowns are covered across the AI Profit Wire signal archive. The dreaming feature accelerates this compression because accumulated codebase context reduces the per-task cost of every subsequent job. The practical risk for any operator paying standard mid-level rates for iterative coding work is that the market rate for that work is being reset by a tool their developers are already using on the clock.
What’s the move on Claude Code for operators?
The move is to identify the most frequently repeated coding tasks in the current workflow and test whether Claude Code can handle them before the next development invoice arrives. Boris Cherny, who heads Claude Code at Anthropic, stated the strategic direction plainly at the event: “The default isn’t ‘I’m going to prompt Claude,’ the default is now ‘I’m going to have Claude prompt itself.'” That framing matters because competitive advantage has shifted from knowing how to use the tool to knowing what to build with it. The dreaming feature makes iterative, repeated work on the same codebase the most financially attractive entry point, since accumulated context reduces cost with every subsequent run. Operators who continue treating Claude Code as a developer productivity add-on are working from the wrong mental model, because the companies presenting at this event were using it as the developer, not a support tool.
Source: MIT Technology Review
