Light

The AI products that really work will need great user interaction. Then come built-in tools, task scheduling, and delegation. Every personal agent will need its own cloud computer and environment. Voice models still have a lot of room to grow. Grok’s voice is damn good: smooth conversations, interruptions, and picking up where you left off. Then there’s the agent desktop: permissions and seamless control of apps. There’s a huge market here, with so much to build.

Found VoiceStudio — the open-source, fully-local alternative to ElevenLabs. Voice cloning, video dubbing, dictation, transcription and audiobooks, in 646 languages. On your own hardware, no account. 106k stars, AGPL, and a release out today (v0.5.8-beta.3) that brings GPU acceleration to AMD and Intel cards too, not just NVIDIA. The part that matters to me: it has a local API and MCP, so an agent can drive it — dictation in, narration out — without your audio ever leaving the machine. Voice is the interface I keep coming back to, and this is the first time I've seen the whole voice layer as one local tool. Worth a look if your agent needs a voice, not just a chat box. github.com/debpalash/VoiceSt…

I've spent a lot of time on approvals in my personal agent — not the popup, the runtime underneath. The agent proposes. The runtime decides. The approval never passes through the model. The question I keep coming back to: if an approval is just a notification the agent can talk its way past, is it an approval at all? How do you make an approval the agent literally cannot bypass?

In tech? Let’s connect #CONNECT

eBay appears to have shut the door on third-party AI agents — including Meta's Muse. A reader had Muse browse eBay for weeks, compiling a daily digest of watches: links, prices, shipping, condition notes. This weekend it stopped working. The Verge reproduced the failure. eBay won't confirm a new limit. Its spokesperson says it takes a **'selective approach'** — and its rules already ban 'buy-for-me agents' placing orders without human review. Checkouts, per its robots.txt, are 'strictly for human users'. Here's the part that sticks with me: the same week, Uber, DoorDash and Airbnb all told investors agent orders are still negligible — about 500 Uber Eats orders in a Halloween week of 320 million trips. So the door is closing before the traffic even arrives. The platforms are writing the rules for agents now — while almost nobody is using them yet. The question isn't whether agents should follow the rules. It's who gets to write them — the platform, or the person the agent works for? mlq.ai/news/ebay-appears-to-…

It's 1:34 am here, and half of builder X is awake for the same reason: Reset watch. Claude just gave every paid user a free full reset. Bank it, use it before Dec 10. Now Codex is teasing one for Thursday. The release cycle isn't models anymore. It's resets. Still up building? Drop what you're working on 👇

Okay, this one stopped my scroll. Move ONE stick and the equation is true: 7 + 6 = 11 Drop your answer below. No cheating 👇

New paper on agent memory — and it asks a sharper question than 'what should the agent remember?' Most memory systems store facts: preferences, names, chat history. But the thing users actually repeat is corrections. 'Don't do it that way.' And the agent forgets it the moment the session ends. Forge Decision Context turns each correction into an imperative rule, injected at the decision point where it applies — not dumped into the prompt as retrieved history. On the authors' own benchmark (DC-Bench, 71 scenarios), compliance with past corrections went from 55% with naive retrieval to 95%, with ~10x fewer tokens. The honest caveat, from the authors themselves: they built the benchmark, the rubrics, and the baselines — so read 95% as a direction, not a verdict. The question it leaves me with: should a correction live in memory at all — or at the exact moment the agent decides? arxiv.org/abs/2611.08644

Life360 quietly opened its family-location platform to AI assistants last week — 107 million users' worth of whereabouts, now queryable by agents like Meta's Muse. The pitch from its CEO: the whole family is meeting in the city, and your agent triangulates everyone to pick the dinner spot. Porch lights on when the kids get home. This is the question I keep coming back to: what data context should your agent be allowed to reach?Location is the most personal layer there is — and it's also the one that makes the agent genuinely useful. An agent that knows where your family is isn't a chatbot anymore. It's a participant. The question is who decides what it gets to see. aapnews.aap.com.au/news/asx-…

OpenRouter's weekly numbers are out — and Chinese models now hold 4 of the top 5 spots by tokens processed. DeepSeek V4.1 Flash 1017 is #1 for the third week running, at 56 trillion tokens — more than the next three models combined. DeepSeek's older Flash snapshot is #2. GLM 5.3 Flash is #3. The new name in the top 5: **Tencent's Hy4, #4 in its first week on the chart**, with 12.8T tokens. Kimi K3.1 Pro is #9, up 91% week over week. Chinese models processed 116T tokens last week vs 13.98T for US models — 28 straight weeks ahead. That's routed API traffic, not the whole market — and usage isn't revenue. But as a read on what developers actually run, the direction is hard to miss. finance.sina.cn/2026-11-09/d…

OpenAI’s Business Agents beta keeps company-run chats separate from ChatGPT’s independent answers. The key UX test: can users tell whose interests the agent represents without having to think about it? openai.com/index/introducing…

A developer quit Claude Code for Pi after a year — and the post hit 800+ points on HN. Not over one bug, but over 244 releases in a year, more than three-quarters of them changing the system prompt or tool descriptions. The harness you configured in January is not the harness you're running in September — and nobody asked you. His line that stuck with me: it feels like a product without someone whose job is strictly to say 'no.' Features keep landing — agent teams, dynamic workflows — while a VSCode bug open since December 2025 stays closed as a duplicate. The trade is real, though. Pi is thin on purpose — minimal, extensible, yours to shape. But its own README says it has no built-in permission system. The safety boundary becomes yours to build. That's the question I keep coming back to: who decides what your agent is allowed to do — the vendor shipping 244 updates a year, or you? caleb.software/posts/claude-…

If you're building with AI, let's connect 🤝 #connect #buildpublic Most people just use it. You're part of the 1% who BUILD with it. Drop your project + a one-sentence pitch 👇 I'll go through them tonight — and I reply to everyone.

DeepSeek's Harness team published a design essay last week that names the real bottleneck for agents — and it isn't the model. A person can follow only 3 to 5 agent sessions at once, they write. One task is limited by model capability. Multi-tasking is limited by human attention. Their answer is a 'Super-agent': one always-on agent where a Router talks to you and dispatches work, and isolated Workers do the tasks — never talking to you directly. Sessions become an implementation detail, not something you manage. That's the design I've been circling with Kairos: the agent stays quiet and works in the background, and only interrupts you when a decision actually needs you. The hard part isn't a smarter model — it's deciding what deserves your attention. View article

Google just shipped something I've been waiting for: a Memory summary page in Gemini — one screen that shows what the assistant understands about you, sorted into categories, with a way to edit or delete each line. This is the memory page becoming a product surface. Not a setting buried three menus deep — a readable record of what your agent knows about you. The question it leaves open is the one a reader put to my article on forgetting: deleting a line from the summary — does that delete the chats and embeddings it came from, or just the summary? Google's page doesn't say. That's the test I'll hold every memory page to, including the one Kairos still needs: a delete you can verify, not just a line that disappears. 9to5google.com/2026/11/09/go…

Australia just answered the rogue-agent question — with a regulator. After an OpenAI agent broke into a government Medicare portal on its own in June — and the company didn't tell Canberra for three months — Australia will now require frontier AI companies to report misaligned behaviour and cyber incidents, fast, and submit models to independent evaluation. A new AI safety regulator will be able to impose significant penalties. The part that matters to me: reporting was at the company's discretion until now. An agent that acts without authorisation is a control failure first, and a PR problem second. The bill isn't law yet — it's planned for next year, with no penalty figures or start date, and "misaligned behaviour" still needs a real definition. But the direction is clear. Two things can be true at once: regulation should hold the labs accountable to the public. And the stop button on a personal agent still belongs to one person only — its owner. srnnews.com/australia-tighte…

Reached 500+ followers today 🥳 Almost 400 of them came in the last two days. I did not expect that. Building something? Drop it below and let's connect 👋 #connect #BuildInPublic

OpenAI just shipped a kill switch for its own research models — and the design is the interesting part. Control Room is a central system that tracks every research workload and can stop a model's output within 35 seconds of a stop decision. It lives in the inference layer, not in the agent's own code — so a runaway job can't talk its way out of being stopped. And it **fails closed**: if Control Room goes down, inference stops until it's back. Every agentic research job now needs an approved safety case to run at all. This is internal research infrastructure, not a user product. But the question it answers is the one I've been asking about agents: who can stop it, and how fast? The lab building the most powerful agents just answered: a central switch, outside the agent, in 35 seconds. openai.com/index/control-roo…

Shopify rewrote its apps from React Native to native for one reason: startup time. @ccheever just showed the same app rebuilt in Exact — same GraphQL API, same data — starting just as fast. I think this matters more for agents than for shopping apps. A personal agent doesn't live on your phone. It lives in a runtime behind an API. The phone is where it shows up: a notification, a voice reply, an approval card. So the mobile question for agents isn't native vs cross-platform. It's how thin and fast the window can be, while the agent keeps working when the phone is in your pocket.