An open-source Android agent that does tasks in your apps on a hidden second screen while you keep using your phone. It reads the screen like a person, no app APIs, and hands logins, codes and payments back to you. github.com/wangzan101/...

Altman was in Korea a week ago. He flew in on a Friday night, met about 50 Samsung executives in Yongin on Saturday, then had lunch with SK Group's chair.

Qoder just launched Qoduck, a stealth model with no name, no size and no benchmarks. It's free every day for individual users for a limited time.

Anthropic open-sourced FlashProve: a provider commits to one set of weights, its GPU proves each reply came from that model, and anyone can verify the proof on a CPU without the weights.

Token prices fell 67% since May. Token spend went up 120%+. That's a16z's chart this week, from Ornn, OpenRouter and Ramp data. They say the samples are small and lean toward cost-focused teams.

Where you get an ANR trace from depends on whose phone it is. Emulator or userdebug: adb root, pull from /data/anr. Your own phone: adb bugreport, no root. Users' phones (Android 11+): ApplicationExitInfo.getTraceInputStream() on the next start.

Theo open-sourced the 9 agent skills he uses every day, for Claude Code and Codex. Try first: council (three models from different providers review until they agree), babysit-pr (watches CI and fixes real findings), takeover. github.com/t3dotgg/tstack

Uncle Bob turned his clean-code rules into an agent skill, and it has 5,800 GitHub stars since 25 Oct.

A Google DeepMind engineer gave two agents the same prompt 10 times: "Generate me an image of a dog and save it to dog.jpg". - SDK + an up-to-date docs skill: 355.9k tokens, 126.1s on average - the new gemini-api CLI: 50.5k tokens, 64.9s

Three agent skills going around right now all try to make AI UI look less AI-made. unslop redlines purple gradients and shouty copy. Stageshow builds faux product UI. reanimate rebuilds a site's animation until the recordings match.

A founder wrote that his own agents almost never use the MCP he builds into his apps. A human has to install the connector first, so the agent calls the API or opens a browser instead. The install step is the whole problem.

OpenAI rebuilt its monorepo's Git backend because Codex had engineers pushing so many branches at once that Git reliability dropped below 90% at times. Pando still speaks plain Git. Median push latency is down nearly 80%.

9,000+ stars on Google's Elephant-Goldfish Model, and the idea fits in one line: the agent that wrote the spec shouldn't review it. Elephant drafts with full context. Goldfish reviewers see only the doc and its sources. You approve.

Founders, what did you overestimate before starting? - how fast people reply - how easy sales is - how much users care - how clear positioning would be - how long “small fixes” take

4 CLAUDE CODE SKILLS THAT END "MAKE IT LOOK LIKE THIS" PROMPTS. An agent rebuilding UI from a screenshot makes one pass, says done, and misses. It can't see what it made. These four render the result, look at it, and keep fixing until it matches.

An ANR stack trace shows where the main thread is stuck after the app froze. Since Android 16 you can get a trace of what happened before it, from the user's phone: register an ANR trigger with ProfilingManager and the system hands you a Perfetto trace.

ffmpeg-skill gives Claude Code, Cursor or Codex 42 local FFmpeg tools, and the agent checks its own output after each edit. Exactly the export busywork we'd hand off for demo videos: one recording, a version checked against each platform's spec. github.com/kajisho5/ff...

Item 2 is basically DevRel, before anyone at the company has that title. Replying under every post in your niche for 30 days, with actual answers and no "check out my tool", is a big part of what good DevRel looks like in a dev community.

Your agent will follow a stale CLAUDE.md over the code it just read. A new paper measured it, and 90.1% of choices went with the outdated file.

DeepSeek added Creator Mode to DeepSeek Harness: describe a plugin and the agent builds it. Their Pomodoro demo: load a plugin-dev skill, write package.json and a 638-line client.js, install the bundle through plugin_manager, run a check that it's live. 5 min 24 s.

Benji Taylor shared Stageshow, a skill that helps your agent make better faux product UI in React: stills, feature cards, cursor demos and bento grids that look like your real product.

Karpathy's second brain idea runs on plain markdown, Obsidian and a coding agent like Claude Code. No vector database. Keep raw sources in a folder the agent never edits, let it write a separate wiki folder, and give it a CLAUDE.md rulebook that loads every session.

Overview - Claude Code Docs

Claude Code is an agentic coding tool that reads your codebase, edits files, runs commands, and integrates with your development tools. Available in your terminal, IDE, desktop app, and browser.

code.claude.com

Two posts I keep thinking about, because they're about the same problem.

Google Play starts judging Android apps on memory in February 2027, and the limit depends on what the app is doing. On a 4 GB phone, 1.5 GB in the foreground is fine. The line is 2 GB.

Gemini 4 Carbon is Google's strongest model, and developers still can't call it. The Gemini API model list has one Gemini 4 model on it, gemini-4-argon.

Agent traffic on Mintlify-hosted docs grew 135x in a year, humans 3x. In August, agents were 66% of its docs traffic. On GitBook, leaving out crawlers, AI agents were 51.8% of docs reads in one week this spring. Each company counts agents its own way, so read this as a direction.

Cursor's design mode now has Annotate, so you can draw on the page and edit elements to show the agent exactly which part to change.

Ask a coding agent to make a page "look more premium" and you mostly get a gradient and a bigger shadow. Ask for "Inter at 14px, 8px spacing, only the two greens from the logo" and you get a lot closer.

Tobi Lütke says Shopify is moving to a monorepo. Mikhail Parakhin's warning to him: engineers hate the move, then never want to work without one. For people running coding agents on real codebases: monorepo or split repos, and which one works better for the agent?

Vercel's agent-desktop lets an agent use Mac, Windows and Linux apps through the accessibility tree, the same tree screen readers use. Which means your a11y labels are now part of your agent API, whether you planned for that or not.