Quick Hits

• The day's sharpest argument: @poteto makes the case that PR volume matters for ordinary engineers now that agents can turn anyone into a "coding machine," directly answering @emilkowalski's skeptical question about why anyone needs hundreds of PRs a day; @larsencc sums up the mood by declaring "classic software engineering is dead."
• Agent interoperability moved on two fronts: @btaylor published the first draft of Sierra's Personal Agent Protocol with 35 new design partners, and @Dan_Jeffries1 says hundreds of harnesses can now talk, group chat, and coordinate for free, a "major unlock."
• Model watch: @0x0SojalSec relays leaks that Qwen 4's 27B variants could ship this week, and @xvremyaperemenx notes Claude Opus 5.5 Fast costs $40 per million output tokens versus $60 for GPT-6.1 Sol Ultrafast while winning 15 of 17 Artificial Analysis tests.
• ChatGPT dots can now start work in Codex, follow existing threads, and edit Scheduled Tasks in ChatGPT Work, per @thsottiaux.
• A caution from @NathanFlurry as Code Mode discussion spreads: QuickJS and WebAssembly sandboxes are "not secure by default," and he lists the specific failure modes to check.

The Volume Debate: @poteto Says PR Counts Matter Again

The bookmarked debate of the day is whether shipping volume is a legitimate metric in the agent era, and @poteto's answer is a qualified yes. Responding to @emilkowalski, who asked "how shipping 100s PRs a day makes sense," poteto concedes that volume used to matter only at the tail ends of a large team, the struggling performer on one end and the "coding machine" archetype on the other, with impact doing the real work. But with frontier models writing code faster than any human, "everyone now has the ability to become a coding machine," and poteto argues that if your PR count hasn't moved, you should wonder why. The prerequisite is trust: without putting in the work to trust your agent's output, you can't scale. On cost, poteto reframes token spend as "cost per intelligence," predicting that what looks exorbitant today will be affordable in six months to a year, with a single engineer plus agents doing the work of tens or hundreds. The job itself has changed: the engineer is now "the machine that writes the software," running what poteto calls a Michelin kitchen.

The reactions piled on. @larsencc quotes poteto's follow-up on "skill compression," the iPhone analogy where craft shifts from mastering equipment to using it creatively, and predicts a generation of builders who never write a line of code. @JoshARosen posted "Software Factories for Skeptics," aimed at the AI-pilled who still doubt factories, and @hwchase17 amplified @caspar_br's counterpitch: "Or don't pay, and use our open-source software factory instead." @mark_k relays @PatrickToulme's recap of an NYC AI dinner, unverified chatter in which attendees claimed most people run agents writing, reviewing, and submitting PRs 24/7, code review is "not super productive anymore," and performance engineers at inference providers now mostly manage agent loops.

Agents Learn to Talk, Get a Protocol, and Gain Hands

Interoperability is suddenly crowded. @Dan_Jeffries1, quoting @HermanNarula, argues people aren't grokking that hundreds of harnesses, personal and public, including Muse, Grok Bots, Dots, Claws, Hermes, Nautilo, and "the one you vibe coded," can now talk, group chat, thread, share rooms, and coordinate for free, with swarms able to borrow each other's compute. On the standardization side, @btaylor announced the first published draft of Sierra's Personal Agent Protocol, launched Tuesday with Meta, Genesys, Instinct, Rocket, Shopify, Stripe, and Walmart, plus 35 new design partners, with the stated goals of consumer choice and letting companies know when they're dealing with an agent. @tweetsbycolin finds the protocol, which he calls Poppy, more pragmatic than MCP because web.browser_session_endpoint allows authenticated AI usage without rebuilding your service as MCP tools, while flagging the gap: there's no way for operators to tell services who is signed in, and he suspects many services would simply trust the big providers. Meanwhile @thsottiaux flags a dots upgrade that lets your dot start work in Codex, follow up on threads, draw on ChatGPT conversations and automations, and review and edit Scheduled Tasks in ChatGPT Work.

Model Watch: A Qwen Leak, Price Math, and an SI Rebrand

@0x0SojalSec relays that Qwen 4 is coming and "beats GPT-6, Fable 5," with 27B variants landing first, early samples allegedly approaching Fable/Opus-level quality, a target window of possibly this week, and open-source October getting interesting after a quiet stretch; @QwenDevs teased only "big or small?" On pricing, @xvremyaperemenx does the math: Claude Opus 5.5 Fast at $40 per million output tokens undercuts GPT-6.1 Sol Ultrafast at $60, and Opus won 15 of 17 tests in the Artificial Analysis comparison. @morganlinton retracts his "RIP Jev" take and is building a decision-model eval suite for @VulcanBench, citing @CompleteSkeptic's launch recap: trillions of tokens served daily, 29.4% of the Fortune 500 showing up, and a big Series A from a16z, all within three weeks. @Cloudflare expanded its Clef family with Clef-omni, natively processing audio, video, images, and text in one pipeline, cut Clef-flash pricing, and boosted Clef inference speeds by up to 2.0x. In branding news, @Benioff declared "the era of Super Intelligence is here" and rebranded AIForce as SIForce, a claim @elonmusk met with satire about moving all SI testing to Delta flights where internet access is impossible. The same @PatrickToulme dinner recap predicted Fable 5.5 will be stunning but that OpenAI's December Bel release will beat it.

Toolchain Notes: Sandbox Security, Fast Loops, and New Tools

@NathanFlurry's security reminder is the most actionable toolchain item: V8 isolates and QuickJS are not secure by default, and his checklist includes same-process isolates leaking API keys via one CVE, Spectre through Date.now, SSRF via guest fetch, botnets from unwhitelisted egress, and unkillable hot loops, with his agentOS Secure Exec answering via sidecar processes, a virtual OS, and deny-by-default egress. For iteration speed, @dillon_mulroy endorses @0xblacklight's take that @alchemy_run plus @Cloudflare lets you skip the dev server and deploy straight to the cloud, a loop he calls great for agents. @dhh showed Gliff, an app for reaching remote Omarchy boxes with tabbed servers streaming over SSH with GPU acceleration, coming in RS 4.5. @MatthewBerman's open-source "Reverse Engineer Anything" project hit #1 on GitHub trending, captioned "Software is done." And @lucasmeijer shipped a free coding tool built around three constraints: no lock-in to one AI lab, cloud agents that can't touch his laptop, and staying connected to code he no longer writes by hand.

Code That Builds Actual Machines

Beyond software, @aakashgupta details @jeffholden's emergence from six years of stealth with $250M and Atomic Machines, whose Matter Compiler builds working micro-machines from code alone with no per-product tooling. The first product is the strategic tell: PrimeSwitch, a 150-amp relay for 800-volt AI data centers that opens in roughly 50 microseconds, about 1,000x faster than a normal contactor, aimed at the one market where demand is effectively infinite. And @Scobleizer spotlights @keenanisalive, a Carnegie Mellon professor whose LLM-generated dragon is 27kb of GLSL defining a closed-form implicit surface, not a mesh, NeRF, 3DGS, or video model.

Practical Takeaway

If you take one thing from @poteto's argument, make it the precondition rather than the metric: volume only becomes a meaningful signal once you've built genuine trust in your agent's output. Before chasing PR counts, instrument the trust itself, track what fraction of agent-written PRs you accept unmodified, how often review catches real defects, and where verification bottlenecks your loop, and treat that trust rate as your scaling constraint. If you're running JS or WebAssembly execution for those agents, audit your sandbox against @NathanFlurry's list first, because a fast factory on an insecure foundation is a liability, not a productivity win.