You post more. Reply more. Keep checking the numbers. Why does getting seen on X still feel like a guessing game?

Today’s architecture article looks at how those three fit together: what each one records, when a result can safely be called saved, and what recovery should do if the app crashes halfway through.

I use Amphetamine to keep my Mac awake while Claude Code or Codex keeps working. It’s a free little menu bar app. It kept my Mac running for six days straight. Even with that much heavy use, I only just managed to use up the allowance from all five GLM-5.3 resets before they expired. The embarrassing part is remembering that I specifically asked the Mac NOT to sleep. Close the lid and put it in a bag without ending the session, and suddenly I’m wondering why my bag is getting warm. Useful on the desk. Before it goes in the bag, I make sure it’s actually asleep.

Zed’s Delta puts handoff inside Claude Code: share the conversation and uncommitted changes, then a teammate joins from their own session. That plugin approach makes sense. You can hand work over from the tool you’re already using. We’ve taken a different approach with OwlRunKit, our own project-local tool. What I find useful is having the current goal, work items, owner, verification evidence and next action recorded outside any one conversation. A new session doesn’t have to guess which result is current or what still needs doing. Delta is worth a look if you want shared coding sessions and file sync. If you want a lightweight, project-local way to carry task state between agents and sessions, I’d recommend trying OwlRunKit. You don’t have to make every task a formal process; the verification level can match the work. The comparison below uses Delta’s official documentation and our experience with OwlRunKit. I haven’t tested Delta, and OwlRunKit isn’t a replacement for its live collaboration features.

The reset is here, and @thsottiaux is back to romancing @poteto. Between the love posts, could we get the remaining Codex allowance right next to the model picker? The usage button has moved out of the menu and into the sidebar. How many tries did it take everyone to find it this time? I’m already choosing a model and typing my next task. Just show me what I have left right there.

GLM-5.5 feels close. All five weekly limit resets I received after the source-code leak expire today. Looking forward to the release. Rumor has it that it’s on par with Claude Fable. Curious to see how it actually performs.

An agent makes it very easy to say yes to a small task. Fix that old bug. Try another layout. Add the feature that never quite made the backlog. Orestis Ioannou wrote about doing more of this with coding agents, then spending more attention on their questions, plans and results. He likes what the tools let him build. He also says he is tired. What interests me is the work that wasn’t going to happen at all. Making it cheaper doesn’t automatically make it more important. Before starting another agent, I’d want to ask: if this still took an afternoon, would I care enough to do it? Sometimes the honest answer should stay no. oioannou.com/blog/ai-promise…

$1,200 a month for Claude: fine. $144 for code review: let’s discuss it for 30 minutes. That’s how one developer described a budget discussion in their six-person team. Their lead eventually pulled up three past incidents and asked which one they would have paid $144 to avoid. But the part that caught my eye was what they cut: a planning tool nobody opened, an autocomplete subscription kept out of habit, and an expensive AI plan sitting idle most of the week. I wouldn’t assume the review tool was worth it just because it was cheap. But before arguing over that $144, I’d want to know why the unused subscriptions had survived so long. reddit.com/r/ClaudeCode/comm…

I don’t always need AI to make me a new presentation. Sometimes I just need it to fix the one I already have. Apple says its November Creator Studio update will let external AI agents read and edit Keynote, Pages, and Numbers through MCP. Keynote is also getting an open, JSON-based .keyx format that external tools can work with. The scenario I care about is pretty ordinary: change a few slides, keep the parts I’ve already reviewed, and leave me with a deck I can still edit. Please don’t make me rebuild everything around a new export. This is an announcement, not something I’ve tested. MCP support requires macOS 27. But giving agents a direct way to work on existing documents is a much more interesting direction to me than another “generate a deck” demo. apple.com/newsroom/2026/10/a…

Your AI says “working.” What proves it? A closer look at how the interface, coordinator, and workers turn activity into verifiable progress.

An AI agent is using your computer. Can you still use it too? One developer described fighting his agent for Chrome’s mouse and focus while trying to code. He moved it to a separate cloud desktop so he could get back to work. That’s the part I care about with computer-use agents. Can I answer a message or open another tab without interrupting them? Can I step in when something goes wrong, then let them continue? If I have to sit there with my hands off the keyboard, the agent is working, but I’m still waiting.

Codex makes it way too easy to start one more task. One more worktree. Another set of dependencies. More build files. 133 worktrees and roughly 90 GB on my Mac later… Now I need an agent to clean up after the agents.

For everyday local AI, I’d still pick a Mac. I use mine for local models and the rest of my work. That convenience matters. I don’t want another machine to manage just to ask a model a question or try something out. A dedicated Linux box can make sense for a specific workload. But “macOS is bad for local AI” misses why people choose it: they want to get things done without turning their setup into another project. If the concern is an agent accessing personal files, let’s talk about its permissions. That deserves a proper discussion. It doesn’t erase the appeal of a computer that’s easy to live with.

Switching things up today. Coffee can wait. Let’s steep some of this first. Getting older, gotta look after myself 😂

Terra 6.1 feels like a really good fit for dots. I added it to my pelican SVG test at max effort. Not the prettiest bird in the lineup, but the speed caught my attention. Then I gave it a small assistant task: browse the official docs, follow a source link, check saved files, and write up the findings. About 2 minutes 15 seconds, including tool waits. That’s the work I keep giving my assistant: look something up, check it, save the result, tell me what happened. Less waiting makes a real difference. Early days, but Terra 6.1 is my default for new assistant tasks now. The grid, A–I (labels, not rankings): A: GPT-6 Astra · max B: GPT-6 Luna · max C: Main assistant session · model unverified, existing conversation context D: GPT-6.1 Sol · max E: GPT-6.1 Sol · high F: GPT-6 Astra · low G: Grok 4.7 · xhigh H: GLM 5.3 · max I: GPT-6.1 Terra · max These are the selected or reported model settings. C wasn’t a fresh-context run, so this is a visual comparison, not a controlled leaderboard.

Congratulations to Harry Kane on winning the 2026 Ballon d’Or! Years of hard work and consistency, and now this moment. So well deserved. 👏🏆

Your AI app updates successfully. Can yesterday’s unfinished work still continue? This article explores how to upgrade an agent runtime without losing reviewed work, stranding active tasks, or mistaking a software rollback for data recovery.

Could we start sharing robot behaviors the way we share small software projects? Microduck’s public training repo includes walking and getting back up after a fall, with instructions for exporting and sharing trained policies. Those are the models that control how the robot moves. For a student or developer, that suggests a concrete project: change one behavior in simulation, try it on the robot, and compare what breaks. Much more useful than watching the same polished routine again. There is still real work involved. Training requires a CUDA GPU, and a vendor’s sim-to-real video is not an independent reliability test. I’d like to see two people train the same get-up motion and compare what changes when it reaches a real floor. That would be a robotics demo worth discussing.

Tibo’s already reset the limits, worried we might go over. Somehow generous and stingy at the same time. It’s kind of endearing. @poteto

Your client is waiting. The meeting-room TV is recommending cartoons. That seems like a better starting point for an AI coding project than “build me something impressive.” Khazix describes using Codex to redesign the home screen on his company’s meeting-room TVs. Screen sharing, input selection and settings stay easy to reach. The entertainment interface is still available, just no longer the first thing people see. I like the modest ambition here. Nobody needed the TV to become an autonomous coworker. They needed it to behave like a meeting-room screen, with the controls people actually came to use.

GLM-5.3 has quietly given me a problem I didn't expect: more quota resets than I can use. Z.ai made a serious mess of things with the ZCode code-upload controversy. It subsequently open-sourced the client. Since then, it feels like there's been less noise and more attention to the service. I'm a heavy daily user. It feels faster now, too. Better infrastructure? Less traffic? I don't know. What I do know is that I still have 7 weekly resets and 4 five-hour resets banked. I genuinely can't get through them all. That doesn't erase the earlier mistake. But this is the kind of improvement I actually notice as a customer. I hope Kimi takes something from this. The quota burn already felt excessive to me back in K2.7. Now K3.1 Pro's launch has also been caught up in allegations about undisclosed paid promotion. Those allegations still need answers, not a verdict based on screenshots. I'd much rather see that energy go into listening to paying users, addressing their complaints and making the service better. Give people a reason to recommend it themselves. Even OpenAI is out here handing out quota resets. Nobody gets to coast on a model launch anymore. Less theatre. Better AI. Treat your users well.