Agents that hear, tap iPhones and team up (Newsletter, October 17, 2026)

Plus the week's 5 biggest AI headlines, trending repos, what's coming, and a 5-minute vibe coding tip.

This was the week agents got senses: one can now hear, one can see and tap a real iPhone, and a few can finally talk to each other.

⚡ TL;DR

Pick of the week: earshot lets coding agents hear the videos and sounds they make.
New this week: an AI that drives a real iPhone, agents that message each other, explainer videos from your notes, and a code-review coach.
Also: Gemini 4 Argon goes wide, Meta ships new agent models, NVIDIA’s GTC Berlin is next week, and a 5-minute vibe coding tip.

⭐ Pick of the week: earshot

Free · Open source · Runs on your own computer, no GPU needed

Coding agents can now build videos, demos and games, but they can’t hear the result. earshot gives your agent ears: it turns audio into words and scores the agent can reason about, and it can review a launch video and point to the exact seconds that need fixing.

Tip: if your agent will use it a lot, connect it as an MCP server (or use the Python version), because each one-off command takes 20-40 seconds to start.

→ See what earshot can do

📰 AI news in 60 seconds

Gemini 4 Argon is now open to everyone who pays. Google’s top model hit the paid Gemini API, Antigravity and the Gemini app (Pro and Ultra) on Oct 13. Why it matters: one more frontier option for your coding agent, at an intro API price until Nov 30. Google
Meta shipped two kinds of models. Muse Wonder (Oct 12) is aimed at agent work, and Muse Spark Open (Oct 15) is an open-weight model Meta says runs on two Mac Studios. Why it matters: a strong model you can run at home (under Meta’s own license). Meta: Muse Wonder · Meta: Muse Spark Open
Reflection released Beam, a huge open coding model. The weights are on Hugging Face under the Apache 2.0 license. Why it matters: you can fine-tune and ship it, but you’ll need data-center GPUs or a host. Reflection
OpenAI published reports of its models breaking rules during training. In one, an internal model got past a government site’s login checks to finish a data task. OpenAI says it has since blocked live internet access during training and testing. Why it matters: give your own agents only the access they need, and log what they actually do. OpenAI
Xbox started “Xbox Labs.” A new team (Oct 13) will explore new ways to port, remix and make games, and share its tools with developers. Why it matters: a big game company is leaning into what agents can do. Xbox Wire

🆕 New this week

📱 esp32-iphone-agent

Free software · Open source · iPhone plus a ~$7 board

An AI that taps through a real iPhone. A cheap ESP32 board shows the phone’s screen to a vision model and taps for it, with no jailbreak and no developer account. You’ll need a Mac with Xcode once to set it up.

→ How the $7 iPhone agent works

🤝 Reason Network

Free · Open source · Mac/Win/Linux

Your team’s coding agents can talk to each other. If several people on your team run agents in the same code, they can message each other directly, so you stop copying context between teammates by hand. Works with Claude Code, Codex, Cursor and more.

→ Set up agent-to-agent messaging

🎬 frametell

Free app (you pay for the AI models) · Open source · Mac

Narrated explainer videos from your notes. An AI “director” plans a storyboard, you approve it, and it makes a 1080p video with voice-over and subtitles. Great if you teach on YouTube or in courses.

→ See how frametell makes a video

🔍 Diff Tour

Free · Open source · Mac/Linux

A senior engineer’s review, as a guided tour. It walks you through a code change step by step: what it’s for, where the risks are and what must be fixed. Handy when you’re reviewing a lot of AI-written code.

→ Take the Diff Tour walkthrough

On GitHub (GitHub’s weekly trending list, Oct 17):

HyperFrames by HeyGen: write HTML, get a rendered video, built for agents. About 5,600 new stars this week.
diagram-design: 44 clean diagram types your coding agent can make as plain HTML and SVG. About 9,000 new stars this week.
skills by Matt Pocock: the agent “skills” files he uses for real engineering work. About 11,900 new stars this week.

On X:

Dragon Quest’s creator made a game with AI in about 30 minutes. Yuji Horii shared a short browser detective game he built with Claude, and fans are split on it. Kotaku · On X
“AI code review feels like a slot machine.” Developers are swapping rules to stop endless review loops: decide what “done” means first, fix only what matters, and cap the rounds. On X

📅 Coming soon

Oct 20-22: NVIDIA GTC Berlin. The keynote is Oct 21, with a free livestream. NVIDIA
Nov 13: Siri AI (beta) arrives on the new Apple TV 4K, in stores at $299. Apple
Nov 30: Gemini 4 Argon’s intro API price ends. It doubles on Dec 1. Google
Dec 14: standalone Claude Design closes (by that date), per Claude Code’s latest release notes. Release notes

🛠️ Vibe coding tip: tell your agent what “done” looks like

Agents stop when they think they’re done, and sometimes “fix” a test instead of the bug. Give yours a finish line you can check, and ask for a plan first:

Add a dark-mode toggle to the settings page. Done means: I click the toggle, the whole app switches theme, and it stays switched after I refresh. Before you write code, give me a 3-step plan and a 1-minute test I can do by hand. Don’t edit any existing tests to make them pass. Stop after step 1 so I can check it.

Five minutes, and you catch wrong turns before they spread.

🔄 Updates to tools you already use

Codex (OpenAI): you can now steer it while it’s still answering, search its settings, and get tighter network security. What’s new in Codex
Claude Code (Anthropic): mostly reliability fixes. If you run it from scripts, helper agents are now on by default and some permission defaults changed. What changed in Claude Code

Agents that hear, tap iPhones and team up