Agents Go Mobile All At Once, Subagents Move to the Background & Spotify's 4,500-Deploy Roast
Agents Go Mobile — All At Once
Two days ago Theo predicted we were "~6 months from most devs moving their code agents off of their laptops." The market apparently heard "this afternoon." June 29 delivered three remote-agent launches in a single news cycle, and the shape of the next interface war snapped into focus: not the editor, but the phone in your pocket talking to machines somewhere else.
Cursor shipped iOS. The headline launch, by a wide margin — Cursor (824 replies, 952 RTs, 10,991 likes, 4.45M views): "Introducing Cursor for iOS. Build from anywhere by launching always-on cloud agents. Or remotely control agents running on your computer from the app. Composer 2.5 is 75% off in the app now through July 5." The two-pronged design — cloud agents and remote control of your laptop's agents — is the same command-center pattern Hamel Husain praised in Codex desktop yesterday, now in a phone app. The replies were a study in 2026 launch friction: a deafening "android wen" chorus (devops_nk, LexnLin, and ancieraav's resigned "it's always iOS first" (58 likes)); EU and Sweden users locked out (testingcatalog, _Hey_Stu); and a small mutiny over the discount that wasn't — Saurav: "Not seeing the 75% discount," capped by Phil Manning: "75% off? It looks like it's a 25% increase in the app." The recurring question was positioning — rich_edelman74: "how is this different from openclaw?" and Gustavo Barbieri: "Like Codex, it should be able to control my laptop, not just cloud agents. Or am I missing something?"
OpenClaw answered within hours — native apps on both platforms, the thing Cursor skipped. OpenClaw (332 replies, 794 RTs, 5,798 likes, 1.47M views): "OpenClaw is now on iOS + Android 🦞 Native mobile apps, finally. Agents in your pocket. Channels, tasks, replies on the go. Run agents from wherever your thumbs are." The timing was not lost on anyone — Francisco Bautista: "After Cursor launched their new shiny iOS app 😅." Reception split hard between delight and bug reports: MrGoober814 was happy ("I've been patient for the Android app and it seems well polished already. Setup over Tailscale with no issues"), while QR pairing fell over for many (kanwisher, victorhaleru, alex_rosenblum) and the frontend took fire — Anon (70 likes): "so obviously made with codex because the frontend is ass." The existential reply came from Merliiin: "what's the point of having it on a phone when any other ai has the same thing? The phone's capabilities will be severely limited, which defeats the main purpose for which openClaw was created."
Theo, meanwhile, was living the thesis. Theo (70 replies, 829 likes, 64K views): "I finally got the T3 Code mobile app and T3 Connect set up. Working on multiple projects across 4 different computers with a single app. This feels like magic." It's React Native (Android in the works, delayed slightly by the earthquakes in Venezuela where the lead mobile dev lives), fully open-source, bring-your-own-subscription. The use-case the replies kept reaching for was the always-on fleet — Malhar Ujawane: "this means I could have a cloud VM as a node and have several agents running 24x7 that I can manage from my phone." Not everyone's sold the experience is there — NickelanddimeO: "For me personally the codex experience is not there yet." The split between camps is shrinking to a detail: not whether the work leaves the laptop, but which app you point at it. (Theo also confirmed he's keynoting AIE — "for some reason.")
Subagents Move to the Background
While the apps fought over where your agent runs, Anthropic shipped a change to how many run at once, unattended. Boris Cherny (453 replies, 307 RTs, 7,578 likes, 451K views): "In the next version of Claude Code: subagents run in the background by default, so you can keep talking to Claude while your subagents work. If you want your agent to run in the foreground, just tell Claude." It's been rolling out gradually to a small subset for a week — which is why half the replies were "wait, I already have this?"
The mechanics matter, and Boris filled them in. On oversight — Cherny: "We forward permission requests from subagents to your main agent, so you see what they're up to. You can also arrow down + enter to 'zoom in' to a subagent and message it." On depth — subagents can spawn subagents "up to 5 layers deep," with a separate-tab UI coming for those who opt into agent teams, and it works with claude -p out of the box. The skepticism was healthy: Harry Read — "so Claude is just going to be doing stuff while I'm not even looking now. what could go wrong" — and two sharp resource concerns. Delali worried about runaway fan-out ("more agents spun up for a simple task… burning more tokens"), which Boris met with a /bug request. And 100YenDev named the unglamorous gotcha: "cache invalidates while waiting on subagents… 5 min cache makes no sense with any subagent work."
The feature people actually want next is steering, not parallelism. LLMJunky, quoting a wish that Claude Code let you control "not just one agent, but a swarm of subagents, mid-conversation": "Mid turn steering remains a top 3 feature they've added over the last year for me personally. It's the only thing I'll miss when we all have 1000tps+." Background subagents make the swarm; the open question is whether you can grab the wheel of the whole swarm while it's moving — Boris's "arrow down and zoom in" is the first answer.
4,500 Deploys a Day — Velocity vs. Value
The day's most-ratio'd thread was a single statistic. ClaudeDevs (210 replies, 139 RTs, 2,595 likes, 974K views), RT'd by Thariq: "Boris sat down with Spotify VP of Engineering Niklas Gustavsson. Spotify ships 4,500 production deploys a day, and 73% of PRs are now AI-assisted." Intended as a flex, it became a referendum on whether shipping velocity means anything at all.
The roast was merciless and, mostly, fair. 0xmediumware (227 likes): "4500 deploys a day to play music? to / play / music." Will Jones (222 likes): "Are they doing a PR for every song?" Gio (213 likes): "Product looks the same for the last 10 years. What are they deploying?" Helder Oliveira: "4,500 deploys to center a div in css." And the line that cut deepest, from Nick Cantelmi: "I'm amazed 27% is AI-unassisted." Several read the number as a quality warning rather than a brag — demonslikeher: "Seeing that number is making me wonder if that's why things that have worked for years have seemingly started to break at random… A lack of quality control due to delegation to AI seems to have entered the production cycle." The cynical-but-plausible take came from Ted Kupolov (10 likes): "The interview is painful to watch. The VP's body language tells you he wants to be elsewhere. Was Spotify forced to do this interview to get the Claude Code contract renewed?"
Pocock supplied the steelman from the other direction. Matt Pocock (82 replies, 25 RTs, 625 likes, 55K views): "Still seeing token anxiety everywhere. We're paying below minimum wage for development and still asking how we can pay less. Don't think in raw tokens, think in value shipped per token." The arithmetic backed him — Sabir Semerkant: "$140/mo… Even the cheapest qualified dev labor is a minimum of $1000/mo. You're paying 8-9x less. Why the token anxiety?" But the best replies pointed at the same hole the Spotify thread did — output volume isn't value. Lynn: "Token cost is visible, but review time is the real tax. Cheap output that creates cleanup work is expensive." Spencer Zhao: "I still want cost per accepted change. Otherwise the agent can look cheap while creating review debt." Joseph Campuzano: execs have "sticker shock… many claim they're much more productive [but] not much actual value is being shipped. Just perceived productivity." And Jens Honack named the actual waste: "not 'the model used 20k tokens.' The waste is spending tokens on vague goals, bad context, missing tests, and loops that don't converge." Between Spotify's deploy counter and the token meter, the lesson rhymes: the number on the dashboard is the cheapest thing to optimize, and the easiest one to fool yourself with.
The Craft — Great Skills, a Personal Wiki & the 6am Rewrite
Underneath the launches and the policy noise, three builders quietly published field notes on the actual work.
Pocock ran a skills masterclass as an AMA. Sidelined from an in-person keynote, he recorded "The Missing Manual: How To Write Great Skills" remotely (49 replies, 210 likes, 39K views) and took questions. The top-five mistakes: "1. Not keeping it DRY. 2. Keeping reference material relevant to only a single branch in the main SKILL.md (instead of putting it behind a context pointer). 3. Allowing no-ops to creep in. 4. Not using leading words. 5. Not using user-invoked skills." His tool for hunting no-ops is the deletion test: "delete the suspicious-looking no-op, then see if behavior changes" — and crucially, "I try not to over-fit to the model." On leading words, the technique is linguistic compression: "words like 'relentlessly' that pack a lot of meaning into a small number of tokens… they help guide AI's behavior." When is a skill done? When "you can no longer find any duplication/sediment/no-ops."
He's also building the environment those skills will run in. Pocock (75 replies, 16 RTs, 956 likes, 92K views): "Doing my first ever experiments with a personal, entirely agent-managed Karpathy-style wiki. X, Discord, Gmail are all being ingested into it every few hours. This is the knowledge base that will serve as the environment for all of my future loops." The implementation is deliberately boring — Sonnet ingesting via the X API, obsidian-compatible but not Obsidian, agentic search and no vector DB — and he's unbothered by scale: 5,000+ markdown files is fine, "just like when a codebase has 5000+ files." Steipete chimed in with his own stack — "birdclaw, discrawl and gog, all backed up to git because you can" — while Travis Hinton supplied the experienced skeptic's caution: "My numerous attempts at fully automated ingestion… have all failed spectacularly. LLMs need to be provided with the frameworks, perspectives, schemas you'd like information organized within — otherwise it's just not useful."
And the human counterpoint to all the loop talk. Thariq (68 replies, 17 RTs, 770 likes, 44.9K views) described his writing process: "do some engineering work, talk to a bunch of people, brainstorm and research with Claude, write a post, give 1 or 2 talks, rewrite the post, give another talk, rewrite the intro, wake up at 6am and rewrite it again, then post." The replies caught exactly why it resonated. Jatin Garg: "Claude can write the draft instantly, and you still rewrote it four times and gave two talks. That's the whole lesson. The model made the writing free, so the value moved entirely to the thinking the rewrites force. You kept the slow part on purpose." Nikhith: "This is why good writing is expensive." Pranav Gajjala: "The rewrite at 6am after sleeping on it is doing more work than the Claude research step." In a week obsessed with deploy counts and token meters, it's the cleanest statement of where value actually pools: in the part you can't delegate.
Also Worth a Look
LlamaParse ships a Retrieval Harness. Jerry Liu (91 likes, 10.3K views): "the 2026 version of RAG over documents… Generalized agents need the right set of tools to scalably search and read through an arbitrary corpus." It exposes four interleavable agent tools — Hybrid Retrieval (agent-tunable vector/keyword alpha), List Files (a scalable ls over an index), File Grep (regex within a file), and File Read (read a subsection past a chunk boundary). The framing is the now-familiar "semantic search alone doesn't cut it; neither does brute-force grep; agents need both," extended from the local filesystem to a managed backend. Nick Venturi's verdict: "maybe now my agents will actually read the pdfs instead of guessing."
Claude lands on Azure, GA. Anthropic: "Claude in Microsoft Foundry is now generally available, hosted on Azure. Azure customers get Claude Opus 4.8 and Claude Haiku 4.5, with Azure authentication, billing, and commitment retirement." Quiet but consequential distribution news — Claude is now a first-class option inside the enterprise cloud where a lot of the budget actually lives.
The open-weights policy fight grinds on. LLMJunky (207 likes, 12.5K views): "Dario is sure doing his best work trying to get open source models outlawed. He's going to keep pushing for this." The thread was mostly heat, but two replies framed the structural argument cleanly — Dan B: "Create the problem, sell the solution." — and Ishu Agrawal: "the biggest threat to Anthropic isn't OpenAI, it's open source models." It's the same regulatory-capture suspicion that's run through every roundup this week, just pointed at a name.
AIE World's Fair takes over SF. swyx's event opened to a stampede — 1,000 registrations on day one alone ("tmr and tuesday going to be absolutely batshit") — kicking off four days that, by his earlier count, are "the epicenter of AI." Matt Pocock had to bow out of his keynote for personal reasons but recorded it remotely; expect the next several roundups to be heavy with dispatches from the floor.