Visual Studio Code 1.127 - The July VS Code release. The headline is agent browser tools going GA - agents can now open pages, take screenshots, and click around to validate their own work, with per-site browser permissions. There's also an overhauled Agents window (session grouping and drag-and-drop), subagent cost transparency, and default terminal sandboxing on macOS and Linux.

Modernize .NET applications in the GitHub Copilot app - Microsoft's .NET upgrade agent now runs in the GitHub Copilot app, with an interactive "upgrade canvas" that tracks assessment, planning, execution, and results in a single dashboard - alongside its existing Visual Studio, VS Code, and Copilot CLI surfaces.

OpenAI had a big launch day on July 9...

Introducing Grok 4.5 - xAI (now branded SpaceXAI) launched Grok 4.5, a model aimed squarely at coding and agentic work, built on its 1.5T-parameter V9 foundation and trained alongside Cursor - the first fruit of the acquisition I covered last edition. It's priced aggressively at $2/$6 per 1M tokens with a claimed 2x token efficiency, and is available in Grok Build, Cursor, and the xAI API (not yet in the EU).

Claude Science - Anthropic launched a beta desktop app (macOS/Linux, on Pro/Max/Team/Enterprise) that pulls research tooling into one place: it connects to 60+ scientific databases, runs code locally or over SSH/HPC, and produces auditable, fact-checked artifacts. It's aimed at researchers, but the local and remote code execution makes it interesting well beyond the lab.

More details on Fable 5's cyber safeguards and jailbreak framework - A follow-up to last week's Fable 5 redeploy: Anthropic detailed the four-tier cyber safety classifier now shipping with it, proposed a public "Cyber Jailbreak Severity" scale (CJS-0 to CJS-4) developed with Glasswing, and opened a HackerOne programme for reported jailbreaks.

OpenAI's new Realtime API models: gpt-realtime-2.1 and gpt-realtime-2.1-mini - Two updated low-latency voice/multimodal models. The full model improves alphanumeric recognition, silence and noise handling, and interruption behaviour; the distilled "mini" tier is faster, with around 25% lower p95 latency across Realtime voice.

GitHub Copilot had a busy week...

GitLost: how researchers tricked GitHub's AI agent into leaking private repos - A sobering one from Noma Security: an unauthenticated attacker could exfiltrate private-repo contents via prompt injection in a public GitHub Issue processed by GitHub Agentic Workflows. Responsibly disclosed to GitHub before publication - and a good reminder of what prompt injection means for agentic tooling.

Claude Code now defaults to "Manual" permission mode - Across releases 2.1.200-2.1.202, the default permission mode switched to Manual on the CLI, VS Code, and JetBrains, and AskUserQuestion dialogs no longer auto-continue by default. There's also a new setting to control how many agents a dynamic workflow spawns.

Claude Code's /doctor is now a full setup checkup - In v2.1.205, /doctor (with a new /checkup alias) went from a read-only diagnostics screen to a checkup that can diagnose and fix: unused skills/MCP servers/plugins versus their context cost, CLAUDE.md deduplication, slow hooks, and permission pre-approvals - reporting what it found and asking before changing anything. Well worth a run!

Announcing TypeScript 7.0 - The long-awaited native (Go-based) compiler port went stable this week, about three weeks after the RC I covered last edition. Full builds are typically 8-12x faster (the VS Code codebase went from 125.7s to 10.6s), editor error-reporting is over 13x faster, and there's a new multithreaded LSP-based language server. It ships as the regular typescript npm package, with a @typescript/typescript6 compatibility package if you need to run 6.x side-by-side - and note the stricter defaults (strict: true, module: esnext, no automatic type discovery).

npm v12 is now GA - The other follow-up from last edition: npm v12 went stable, switching on the security defaults previewed in June - lifecycle install scripts are now off by default (npm approve-scripts), and Git/remote/tarball dependencies need --allow-git / --allow-remote. The same announcement also starts deprecating 2FA-bypass granular access tokens in favour of trusted publishing (OIDC). Worth checking your CI pipelines before upgrading!

Rewriting Bun in Rust - Jarred Sumner announced that Bun has been rewritten from Zig to Rust - by around 64 parallel Claude instances porting 1,448 files (~535k lines) over 11 days and 6,502 commits, with adversarial-review loops checking the work. v1.3.14 is the final Zig release, and the Rust-based v1.4.0 is in canary. The post also quietly disclosed that Anthropic acquired Bun back in December 2025. Zig creator Andrew Kelley posted a considered response that's well worth reading alongside it.

Cloudflare Drop - A neat little tool: drag a folder or zip into the browser at cloudflare.com/drop and get a live static site on Cloudflare's edge in seconds - no account needed. Previews live for an hour unless you claim them into an account.

GitHub Models is being fully retired on July 30 - GitHub has set a hard shutdown date for all of GitHub Models - the playground, model catalogue, inference API, and BYOK endpoints - for every customer, including existing paid ones. There are brownouts on July 16 and 23 before the July 30 cutoff.

A big collection of leaked system prompts - A colleague shared this in our work Slack this week, and it's a good one: a popular (50k+ stars), regularly-updated repo collecting the system prompts behind most of the major AI tools - Claude, ChatGPT, Gemini, Copilot, Cursor and more. Worth a browse if you want to see how these things are actually instructed under the hood, and to pick up a few prompt-engineering ideas along the way. It's community-extracted, so treat it as indicative rather than official.