The AI Autonomous Programming Era: From Code with Claude to a World Where Machines Write Code
The Hand-Raise That Changed Everything
At Anthropic’s Code with Claude event in London on May 19, 2026, engineer Jeremy Hadfield asked the packed room a two-part question that captured the moment software engineering tipped into a new era:
“Who here has shipped a pull request in the last week that was completely written by Claude?”
Nearly half the room raised their hands.
Then the follow-up: “Who here shipped a PR completely written by Claude where they did not read the code at all?”
Nervous laughter. Most hands stayed up.
pie title "Code with Claude — Developer Survey"
"Read & reviewed AI code" : 25
"Shipped AI code without reading" : 45
"Still write code manually" : 30
Illustrative data based on the show-of-hands poll at Code with Claude, London, May 2026.
This was not a fringe group of “vibe coding” enthusiasts. This was a room full of professional software engineers at a flagship industry event, casually admitting that the core practice of their profession — reading and understanding the code they ship — had become optional.
From “Prompt Claude” to “Let Claude Prompt Itself”
The philosophical shift at Anthropic goes deeper than just generating code. Boris Cherny, head of Claude Code, described a fundamental change in how developers interact with AI:
“The default isn’t ‘I’m going to prompt Claude’ — the default is now ‘I’m going to have Claude prompt itself.’”
This is not marketing hyperbole. Anthropic’s engineering philosophy has coalesced around a simple mantra: “Let it cook.”
Ravi Trivedi, an Anthropic engineer, explained the principle: “The key principle is getting out of Claude’s way.” In this new paradigm, human developers should not even see error messages. Claude handles the full loop — generate, test, fail, fix, retest — until everything works.
sequenceDiagram
participant H as Human Developer
participant C as Claude Code
participant R as Runtime/Compiler
H->>C: Describe the task
C->>C: Plan approach (self-prompt)
C->>R: Generate code
R-->>C: Error: type mismatch
C->>C: Analyze error, fix
C->>R: Generate corrected code
R-->>C: Tests pass ✓
C->>C: Self-review & refine
C-->>H: PR ready for review
Note over H,C: Human sees only the final result
The Dreaming Feature
Perhaps the most revealing feature announced at the event is Dreaming — a system where Claude Code agents write notes to themselves, recording useful information about specific tasks. When another coding agent later works on the same codebase, it reads through these notes, spotting patterns and learning from errors previous agents encountered.
This is, in essence, machine learning at the organizational level — agents accumulating institutional knowledge about a codebase and getting progressively better at working on it, without human intermediation.
The Scale of the Shift
The numbers tell a story that would have seemed like science fiction just two years ago:
| Organization | AI Code Claim |
|---|---|
| Anthropic | ”Most software at Anthropic is now written by Claude” |
| Anthropic | ”Claude has written most of the code in Claude Code” |
| OpenAI | Similar claims about internal Codex usage |
| Majority of new code AI-generated or AI-assisted | |
| Microsoft | GitHub Copilot used across majority of engineering teams |
The reflexive loop is striking: Claude Code is largely written by Claude. The tool builds itself.
At the event, Katelyn Lesse (Claude engineering lead) assessed Claude’s current capability: “Claude is probably as good as a midlevel engineer at writing code.” And Angela Jiang (Claude product lead) stated the ultimate goal plainly: “The absolute end state we’re trying to get to is Claude basically being able to build itself.”
The Other Side: Code Without Understanding
Outside the conference walls, the mood is more ambivalent. On Hacker News and Reddit, a growing chorus of developers is raising alarms:
- “The only people I’ve heard saying that generated code is fine are those who don’t read it.” — Hacker News user
- Reports of developers feeling their coding abilities atrophy as they hand more tasks to AI
- Security researchers warning that AI-generated code introduces new classes of vulnerabilities
- Managers exhausted by the volume of AI-generated code they need to review
Lesse acknowledged these concerns while holding firm: “All of the old software development best practices still apply. They’ve applied this entire time. I think there are a lot of people and teams that may have lost sight of them in this moment.”
But she also admitted that Anthropic’s own technical managers struggle to keep up with the code their teams now produce. The productivity gains are real — and so is the oversight gap.
graph TD
A[AI Code Generation] --> B{Code Reviewed?}
B -->|Yes| C[Better code,<br/>faster delivery]
B -->|No| D[Technical debt,<br/>security risks,<br/>skill atrophy]
C --> E[Productivity gains<br/>are sustainable]
D --> F[Short-term speed,<br/>long-term fragility]
style C fill:#51cf66,color:#fff
style F fill:#ff6b6b,color:#fff
Beyond the West: The Singer Who Codes
Perhaps nothing illustrates the democratization of programming better than the story that broke the same week: Chinese singer-songwriter Hu Yanbin (胡彦斌) used Claude to build a community app called “彦火” (YanHuo), now in public beta.
A professional musician — not a programmer — created and shipped a functional application. This is the promise and the disruption of AI programming in one story: the barrier between “can code” and “cannot code” is dissolving.
Codex and the End-to-End Frontier
OpenAI’s Codex represents the other pole of this revolution. Unlike Claude Code’s terminal-based interaction model, Codex operates as an agent-driven simulator — capable of operating a computer, navigating interfaces, debugging, and fixing issues autonomously. A demo video showed Codex completing an entire development cycle without human intervention: reading requirements, writing code, running tests, fixing bugs, and deploying.
The trajectory is clear: 2025 was the year AI learned to write functions. 2026 is the year it learned to build software.
What Happens Next?
Three forces are converging:
- Capability: Models now match mid-level engineers and are improving rapidly
- Autonomy: Self-prompting, dreaming, and end-to-end loops reduce human involvement
- Adoption: From startups to Spotify, organizations are restructuring around AI coding tools
The outcome is not that programmers disappear — but that the nature of programming work changes fundamentally. The most valuable skills shift from writing code to designing systems, reviewing AI output, and defining what should be built.
Whether this produces more robust software or a mountain of unmaintainable technical debt depends on whether the industry can maintain rigorous review practices even as the temptation to “let it cook” grows stronger.
One thing is certain: the hand-raise in that London conference room marked a point of no return.
Sources: MIT Technology Review, OpenAI Codex demo, Reddit/Hacker News discussions