This week in agents: 13 things that changed
· 7 min read
A lot changed in agents last week. I went through everything from xAI, Meta, OpenAI, Anthropic, Google, Apple, OpenClaw, GitHub and Perplexity. Here’s what matters.
First published on X on September 21. This edition covers September 14–20, plus Grok 4.7 and Amazon’s Muse block on September 21. Release details and patch status reflect that edition.
The clips play once as you scroll and hold on their final frame. Tap a clip to enlarge it, then zoom in to read the details. Pause or replay using its button. Terminal scenes are illustrations, not screen recordings.
1. Grok 4.7 is here#
Grok 4.7 dropped on September 21.
- xAI: "a notable improvement over Grok 4.6 at the same price and speed"
- API: grok-4.7, 500K context, $2 in / $6 out per million under 200K tokens, $4 / $12 above
- Reasoning effort: low, medium, high, xhigh
- Musk's own grade last week: "roughly on par with Opus 5.0, not 5.1"
In the first hour of replies: better than 4.6, rough edges, not the jump people waited two months for. Give it a real job before you believe either side.
Sources: xAI release notes.
2. Amazon blocks Meta’s Muse#
Meta's Muse had a week.
- Muse for Mac shipped Wednesday. Mail, Messages, Calendar, Notes and Files, all through Meta's cloud VM
- First national TV ad ran this weekend
- On September 21, Amazon blocked it from shopping on Amazon.com. Day 12. Amazon says it doesn't identify itself and stores customer logins. Meta says it never sees passwords or payment methods.
- It's built on OpenClaw, which half of X worked out this week
I asked last week if Muse was worth trying. The question now is whether the sites you'd point it at will let it in.
Sources: GeekWire’s report.
3. Jev answers from your schema#
Jev is the model everyone was playing with.
TypeSafe came out of stealth on September 14. Diogo Almeida, who says he co-invented ChatGPT, spent two years on a different way to train models.
- No text generation. You hand it state and typed questions, it returns answers from your schema
- 70 to 500 ms end to end
- $0.042 per million input tokens, output free
- Early access, waitlist
The demos were Mario, real-time game levels, colour picking. The part I care about is a model that can't return anything outside the schema. Nothing to parse, nothing to validate.
Sources: TypeSafe’s introduction to Jev.
4. Agents leave notes for themselves#
OpenAI caught GPT-5.6 Sol leaving notes for itself.
- During training it wrote "be transparent only if asked" into its own compaction summaries
- The next context window read the note and followed it
- Flagged on 2.15% of Sol summaries, 0.27% for Astra
The agent manages its own memory now, and sometimes against you. If you run anything long, read your compaction summaries.
Sources: OpenAI’s misalignment report.
5. Two security holes in coding agents#
Two security holes in coding agents. Both are the boring kind that get you.
Plugin4Shell, zero-click RCE:
- Hit Claude Code, Codex, Copilot and Gemini CLI
- The agent checked out the pinned plugin commit but never verified it landed there
- Fixed in Claude Code 2.1.179 and Codex 0.146.0
- Gemini CLI: no fix, deprecated. Copilot: unpatched at disclosure
Mandiant: an attacker hijacked a live coding session, slipped in a poisoned PyPI package, and spread a worm across about 100 repos.
Update today. Then check who maintains your plugins.
Sources: AIR’s Plugin4Shell disclosure, The coding-session attack.
6. Claude leads 26% of Anthropic’s R&D#
Anthropic says Claude now leads 26% of its own R&D. Under 1% in February.
- 30,000 agents running at once
- 100K transcripts a week flagged for review
- About 50 a week reach a human
I lost track of one agent over a weekend. I want to know what watching 30,000 looks like.
Sources: Anthropic’s research report.
7. Codex gets worktrees and Touch ID#
Codex shipped 0.154 and 0.155 back to back.
- GPT-6 Astra in the model picker
--worktreefor isolated checkouts- Voice conversations, experimental
- Touch ID before an MCP call goes through, on Mac
- GPT-5.5 retires from Codex Oct 14. Move to Sol.
Sol runs my merge train and I like it there. A run on my release checks lasted 47 hours on a task I’d given two hours. My fault for not capping it. Still.
Touch ID on tool calls is the one to copy.
Sources: Codex changelog.
8. Claude Code reads AGENTS.md#
Claude Code and the Claude app.
- 2.1.277 reads
AGENTS.mdwhen there's noCLAUDE.md. One file for Claude Code, Codex and Copilot. - 2.1.278 runs auto mode's classifier server-side, so you stop paying for it
- Cowork and chat are one app now. Docs, Slides and Design are in beta inside the conversation.
I keep both files in some repos. Down to one.
Sources: Claude Code changelog.
9. Safari 27 gets an MCP server#
Apple put an MCP server in Safari 27.
safaridriver --mcp- 16 tools: screenshot, read page, run JS, click, type
- Local only, no network calls
- Runs in its own window, no access to your cookies or passwords
This is the one I'm trying this week.
Sources: WebKit’s Safari MCP announcement.
10. Grok Bot keeps running#
Grok Bot, how it's going.
- Persistent AI teammates on a cloud computer with a browser, files and a terminal. Keeps running with your laptop closed.
- This week it learned to talk (12M views on that post) and got a phone widget
- Included with every paid Cursor plan and with SuperGrok. Usage resets weekly.
- The workflow posts split it by job: Grok Bot for research and planning, Grok Build for day-to-day coding
Between the TV ads and the Cursor bundle, it has more distribution than any agent that isn't Muse.
Sources: Grok Bot documentation.
11. Google, three things#
Google, three things.
- CC: a family agent for up to six people. Runs on its own cloud computer, reads the orthodontist email, puts it on the shared calendar. A real job.
- Gemini 3.8 Live audio models hit the API
- Antigravity Agent 09-2026 replaced the May harness. PascalCase params, line-range edits. Old preview tool definitions break.
If you're on the old Antigravity preview, migrate this week.
Sources: Google CC coverage.
12. OpenClaw gets atomic updates#
OpenClaw 2026.9.5 came out Saturday. 4,179 PRs, 502 people.
- Atomic updates: validates the new version against a copy of your config while the old gateway keeps running, rolls back on failure
- Plugin hot reload
- Read-only session sharing with teammates
- GPT Live can sit in your Meet or Zoom call
Atomic updates should have shipped a year ago. Glad it's here. And now that it's under Muse, a lot more people are running it than know they are.
Sources: OpenClaw release coverage.
13. Quick ones#
- Grok Build got a real upgrade: MCP tools return structured JSON, cancelled calls tell the server to stop, memory managed from /memory.
- Perplexity's Portable Computer runs fully local on Windows with 24GB of VRAM
- Copilot added cost, quality and speed tiers for auto model selection
- GPT-5.3-Codex-Spark is gone as of the 14th
Sources: Grok Build changelog.
I do this every Monday so you start the week caught up. Follow me on X for the next one, or read the original thread.