AI that matters, from the architect's desk. Curated and engineered by Saaket Varma, PhD — no hype, just signal.
HOW TO READ THIS Read top to bottom: Claude Code ships, commands pass through a transcript guard, the rm -rf gate blocks destructive ones, safe ones reach v2.1.205.
Claude Code v2.1.205 blocks agents from tampering with session transcript files in auto mode, forces confirmation before rm -rf runs on unresolved variables, and fixes a Windows bug where worktree removal could follow NTFS junctions and delete files outside the worktree. These aren't hypothetical hardening items — they're the precise failure modes that turn a long background agent session into an incident: silent history rewrites, destructive shell expansion, and filesystem escapes. /doctor is also now a full setup checkup that diagnoses and repairs configuration issues rather than just reporting them. If you run unattended or overnight agent sessions, update before the next one and run /doctor as part of the upgrade.
HOW TO READ THIS Read top to bottom: two apps start separate, merge into one desktop app, run on both OSes, and are powered by GPT-5.6.
As of July 9, Codex ships inside the ChatGPT desktop app on macOS and Windows — existing users keep projects, settings, and workflows, and can set Codex as the default view, with Computer Use accelerated by GPT-5.6. The move signals the end of the standalone-agent-app era: OpenAI is betting the coding agent wins by sitting where the other 90% of a user's ChatGPT time already goes, not in a separate window. If your team standardized on standalone Codex, verify the migration preserves your project state — and expect distribution, not capability, to be the competitive axis this quarter.
HOW TO READ THIS Read top to bottom: an agent grows three hands, each reaching into a different Office file format, driving GitHub adoption below.
OfficeCLI (13.3k stars, Apache-2.0, v1.0.134 released July 9) is a single-binary Office suite built for agents: deterministic --json output, stable element paths like /slide[1]/shape[2], a built-in HTML/PNG renderer for visual verification, and one-command skill/MCP install into Claude Code, Cursor, Codex, and Copilot. It closes the read-edit-verify loop on .docx/.xlsx/.pptx with no Office install — the missing piece for agents doing real business-document work rather than demos. If your agents touch enterprise documents, this replaces a pile of brittle python-docx glue with something an agent can actually verify against.
HOW TO READ THIS Read top to bottom: Cursor ships 3.10 with Team MCPs, listed once in an internal marketplace, and every team member converges on shared org-level access.
Cursor 3.10 lets teams publish MCPs to an internal marketplace with organization-group access control, building on 3.9's Customize page that unifies plugins, skills, MCPs, subagents, rules, commands, and hooks at user, team, or workspace level. The real story is governance: agent configuration is becoming a versioned, shared, access-controlled artifact instead of per-developer dotfile sprawl. Even if you're not on Cursor, the pattern — centrally published tooling with scoped access — is worth copying for any org running agents at scale.
Autonomous coding agent packaged three ways — SDK, IDE extension, or CLI — so you embed it where your workflow already lives.
Open-source agentic browser positioned as the self-hostable alternative to ChatGPT Atlas and Perplexity Comet.
Open-source toolkit for building your own AI SRE agents — incident response as code you control.
MCP server giving Claude terminal control, filesystem search, and diff-based file editing on your machine.
Coding-assistant skill that turns any folder of code or SQL schemas into a navigable knowledge graph — works across Claude Code, Codex, Cursor, and Gemini CLI.
SemiAnalysis Anthropic's pre-IPO financials show over $1B in 3Q26 profit — frontier AI economics are no longer hypothetical.