Anthropic dropped Claude Opus 4.7 this week, this model is built for long-running coding tasks, multi-file refactors and complex debugging sessions. From my early testing the model is quite different to what we are used too and it tends to ask lots of followup questions similar to what we've seen with GPT 5.4 and in comparison to Opus 4.6 we see fewer tool errors, and crucially the model now verifies its own outputs before reporting back. Cursor’s CEO called it “a meaningful jump in capabilities but i think it's the model to get cowork capabilities to more people

The vision upgrade is worth noting too. 3x the image resolution of prior Claude models, which unlocks real computer-use workflows — reading dense screenshots, extracting data from complex diagrams, working with pixel-perfect references.

Other new additions:

• an xhigh effort level for finer control over reasoning depth
• /ultrareview for dedicated code review sessions, and auto mode for Max users who want fewer interruptions on long tasks.

One catch for developers migrating, the updated tokenizer means the same input can map to ~1.0-1.35x more tokens depending on content type. Measure before you switch.

The Rise of Agentic IDEs

The code editor stopped being a place where you type and started being a place where you delegate.

Agentic IDEs — tools like Claude Code, Cursor, OpenAI’s Codex, Kiro, and Windsurf — now understand your codebase, make structured multi-file changes, run tests, and open pull requests. They don’t just suggest code. They ship it.

The landscape is crowding fast. Claude Code dominates terminal-native workflows. Cursor remains the go-to VS Code-based option. OpenAI’s Codex just hit GA with cloud sandboxes that let you spin up parallel workstreams. AWS launched Kiro with spec-driven development — less “vibe coding,” more structured intent. And JetBrains released Air, an environment that runs multiple AI agents (Codex, Claude, Gemini) in parallel.

The trend underneath all of this is that developers are moving from working 1:1 with an agent to managing agents in parallel. The bottleneck is shifting from “can the AI write good code” to “can I review and orchestrate what multiple agents produce.”

If you haven’t tried handing a real task, you should.

Patch.md: The Idea That Won’t Leave My Head

Theo (t3.gg) floated a concept in his “A letter to tech CEOs” video that I keep coming back to: patch.md.

The problem it solves is simple. Open-source software is powerful, but the moment you customize it, add a feature, tweak a workflow, change a UI, you’ve forked it. And forks are maintenance nightmares. Every time the original project updates, you’re stuck manually merging your changes back in, resolving conflicts, and hoping nothing breaks.

Patch.md proposes a different approach. Instead of storing your customizations as brittle code diffs, you describe them in plain English — the intent behind each change, not the implementation. “Add a dark mode toggle to the settings page.” “Replace the default auth flow with OAuth2.” “Show billing data in the admin dashboard.

Then, when the upstream project updates, an AI agent reads your patch.md, understands what you wanted, and re-implements those customizations on top of the new version. No merge conflicts. No manual porting. Self-healing forks.

It’s essentially dependency management for customizations. Think package.json, but for the ways you’ve made someone else’s software yours.

The concept is still theoretical, but the pieces are falling into place. With models like Opus 4.7 reliably handling long-running autonomous coding tasks, and agentic IDEs that can branch, test, and commit — the infrastructure to make patch.md real is closer than it sounds.

I’m prototyping something along these lines. More on that soon.

That’s your scoop for this Sunday. See you next week.