Code with Claude London: The Autonomy Dial Just Became the Default

Boris Cherny, who runs Claude Code at Anthropic, said something on the London stage on Monday that I have been waiting six months to hear from a vendor.

"The default isn't 'I'm going to prompt Claude Code.' The default is now 'I'm going to have Claude prompt Claude Code.'"

That is the autonomy dial, stated as Anthropic's default position. The autonomy dial is the framework I use for thinking about how much agency you hand to an AI agent. Low end: you write a tight spec, the agent executes one step, you check, you continue. High end: you set the goal, the agent loops on it without you. Boris just declared the high end is the default.

This is a practitioner take on the Code with Claude London opening keynote from May 19. What the announcements actually mean if you ship code with an agent today, what they get right, and where they are still aspirational.

Last year I read every edit. Now I read PRs.

Cat Wu, who leads Claude Code product, drew the same contrast Boris did. A year ago: "I would read every single edit it tried to make, every permission prompt." Today: "Most of us are running in auto mode to delegate permissions to Claude, and we're checking in after Claude has already tested, verified its changes, and has a PR ready for us."

I made that same shift around February. The trigger was not a better model. It was a better setup. My CLAUDE.md got tight enough that the agent stopped doing the wrong things. My verification hooks got loud enough that I trusted the agent's own checks. After that, sitting in front of every permission prompt felt like co-piloting a self-driving car. More anxiety, no extra safety.

If you are still reading every edit, the model is not your bottleneck. Your scaffolding is.

What Anthropic actually shipped on the autonomy dial

Three things from the keynote that change how the loop runs.

Routines. Claude Code can now run on a schedule, in response to a webhook, or from an API call. Boris framed routines as higher-order prompts. "I'm not the one doing the prompting. I'm the one that creates a routine that does the prompting." If you have ever written a cron job that triggers your agent on every new GitHub issue, you already know this pattern. Anthropic just made it a first-class primitive.

Outcomes. You declare what success looks like, the agent iterates on it. This is the same loop pattern I keep recommending. Define a verifier, let the agent re-run until the verifier says yes. The difference is that Outcomes lives in Claude Managed Agents, so the iteration happens server-side instead of inside your terminal session.

Dreaming. Claude introspects on its previous transcripts and self-improves. Anthropic shipped this as part of Managed Agents. It is the closest a vendor has gotten to the persistent-memory patterns I have been writing about with the Egg protocol and the tulpa model. The shape is right. The depth, we will see.

Two more developer-facing primitives are worth flagging. CI autofix babysits your PR to green: it picks up review comments, fixes lint nits, retries flaky tests, rebases through merge conflicts. Self-hosted sandboxes let Claude Managed Agents execute work on your infrastructure, with first-class support for Vercel, Modal, Cloudflare, and Daytona. Autofix in particular closes the most annoying loop in async coding: the agent ships a PR, CI fails on a flake, the PR sits red for an hour because you are not at your laptop.

I go deeper on the autonomy dial in Chapter 3 of the agentic coding course, "The Human in the Loop." The dial is the single concept that decides whether your agent saves you time or wastes it.

Where the keynote was still aspirational

The other layer of the keynote was direction, not delivery. Three pieces stood out.

Context windows that feel effectively infinite. Lisa Crofoot from the research PM team used that exact phrase. Not "are infinite." "Feel" infinite. The 1M-token context window I already use daily on Claude Code is real. "Effectively infinite" implies long-running agents that hold state across hours of continuous work without losing the thread. That is the bet. It is not the current reality.

Mythos preview. The next model. The demo where Mythos read the entire OpenBSD source tree and surfaced a 27-year-old vulnerability that survived every human reviewer, every fuzzer, every static analyzer thrown at it for three decades. That is a striking result. It is also "preview." Lisa and Boris were careful. Opus 4.7 is the model you ship on today. Mythos justifies the keynote's "the exponential keeps moving" framing. Build for Opus, watch for Mythos.

90% autonomous coding. Oscar Mowen, who leads technology at Mercado Libre, announced a Q3 target of 90% autonomous coding in a fully agent-driven PR loop. 23,000 engineers, 500,000 PRs reviewed, 9,000 apps modernized. That is the most ambitious autonomy-dial setting I have heard a real engineering org commit to publicly. It is also a target, not a result. Worth watching, not worth copying yet.

The practical shift for this week

If you take one thing from the keynote and ship it this week, pick a task you currently kick off by hand and turn it into a routine.

Concrete examples from my own setup, since the keynote landed.

The "triage new GitHub issues" task that used to be a manual gh issue list followed by a Claude Code session. Now a routine. Fires every two hours, opens labeled PRs for the issues it can solve cleanly.

The "rotate the blog cover prompt" task that used to interrupt my morning. Now a routine attached to a PR webhook.

The CI failure on a flaky test. Now CI autofix's problem, not mine.

The point is not the automations. The point is the posture. As long as you are the one running the prompt, you are still the bottleneck the autonomy dial is trying to remove.

Boris's line is the cleanest version of this thesis I have heard from a vendor. The keynote did not give us a new model. It gave us permission to stop treating prompting as the work.

The autonomy dial, the routines pattern, the verification loops that make all of this safe. Start the agentic coding course for the full method.