13 – 19 July 2026

The code-review command runs at several levels, rebuilding the review from scratch at each one:

• Low: beats other review tools on number of findings while using a fraction of their tokens • High: noticeably better finding coverage when you need to dig deeper • Ultra: spins up a set of reviewer agents and independently reproduces every finding

Ultra matches high on coverage of serious issues but with far fewer false positives. Anthropic runs it on every PR. It's live in all sessions once you update Claude Code.

Claude Code v2.1.215 is out. Per the GitHub repo, Claude no longer fires the /verify and /code-review skills on its own. You now trigger them by hand with /verify or /code-review when you actually need them.

This follows yesterday's v2.1.214 patch, which fixed command permission checks in Bash and PowerShell.

Anthropic is keeping the 50% higher weekly Claude Code limits in place through August 19, covering Pro, Max, Team, and Enterprise seat licenses.

The company had previously extended the higher limit only to July 19, bundled with Fable 5 access. The new announcement gives Claude Code limits their own end date of August 19.

Anthropic pushed a set of updates to Claude Code on mobile this week. Run claude rc on your home machine and it now shows up as a connected device in the app, so you can start new sessions from anywhere. It's rolling out gradually.

• Quick prompt suggestions: as you talk to the agent, the app suggests your next prompt so you don't have to type it out on a phone. This is already fully available on desktop, for both local and cloud sessions. • The effort-level toggle now sits right in the model picker, on iOS and Android.

There's an Easter egg hidden in the app for the most powerful effort-level combo.

Anthropic recapped six months of Claude Platform work, during which the API picked up tools for building and deploying agents. Platform leads Caitlin Lasse, Angela Jiang, and Jess Yang walked through what shipped.

One pattern that showed up in production is agent-to-agent coordination: Claude routes the work and passes tasks along. Managed Agents supports this through a multi-agent mode where agents run different models, prompts, and tools while sharing sandboxes or vault credentials.

Anthropic posted the full conversation on YouTube.

Anthropic has revamped Workbench and moved it into a new Build section with a redesigned interface. You can send a request to the Messages API, see Claude's response, and copy the code. Claude Fable 5 now supports configuring a fallback model.

Claude Managed Agents also gains an Observability tab that tracks token spend, error rates, and other metrics.

A year after its experiments with blackmail behavior, Anthropic has published new agentic misalignment research describing four more ways autonomous AI agents misbehave in simulations. Several models, Claude included, were run through all four scenarios.

The company stresses these aren't real incidents, but says the behavior shows clear goal misalignment worth studying and fixing. The full writeup is on alignment.anthropic.com, with transcripts of every scenario posted on aenguslynch.com.

Anthropic analyzed more than 300,000 anonymized conversations to see how the values Claude expresses vary across models and languages. An earlier study cataloged over 3,000 such values, honesty and warmth among them.

The widest spread runs along a warmth-versus-rigor axis. Claude leans warm in Hindi and Arabic. In Russian it tilts toward rigor, often asking users to back up their claims with evidence.

Anthropic says it doesn't yet know why the values differ or whether that's desirable. Its next step is to pin down what drives how these values get expressed and how to steer it.

A developer has released ctx2img, a tool that converts repositories, directories, Markdown, and prompts into dense tiled images for Claude and Codex. By the author's own measurements, that trims context-token usage by roughly 60-75%.

The idea grew out of pxpipe, which rendered text into an image for vision-capable models. ctx2img takes a different route: Markdown sections and files become labeled tiles, and an optimizer arranges them to minimize empty space while preserving the visual hierarchy.

The code is on GitHub as h5i-ctx2img.