We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
🔥 TOP SIGNAL
The practical unit of agentic coding is shifting from a single chat to a supervised handoff. Riley Brown’s CURSOR DELEGATE makes Grokbot the mailbox and Cursor the worker, gives delegated agents a private scratch repository, and returns the cloud agent’s full response; Kody v2026.08.23 makes its factory loop pauseable, scrollable, and inspectable with inline tool calls.
For practitioners, separate the control and handoff surface from execution, and make state and tool activity visible rather than burying them in one chat.
⚡ TRY THIS
Build the mailbox-to-worker bridge. Riley Brown’s posted pattern uses literal triggers such as “Send this to Cursor,” “Ask 5.6 Sol,” and “Delegate this.” Route the request to
gpt-5.6-sol; only setmodel_params.reasoningtohighwhen the user asks for “high” or “5.6 Sol high.” Keep the delegated agent’s scratch work in a private repository and return the full Cursor response, not a paraphrase.Use Codex’s side chat as a second control lane. While the main thread is running, open
/side chatto stage the next question, test review, or follow-up instead of waiting; Derrick Choi calls the feature underrated, and Riley Brown says he uses side chat more than main chat.When the sandbox lacks hardware, move the test—not the goal—to CI. Simon Willison gave Claude Fable 5 this task:
Put https://smolmachines.com through its paces as a fast secure sandbox. Explore what it would take to use this to run untrusted Python and JavaScript code in a way that is limited in what RAM and CPU time it can take up (protection against "while true") with no network access and filesystem access only to designated filesGoal is to be able to use this to execute user-provided tasks for things like data transformationsWhen the web container had no
/dev/kvm, the agent detected the limit, ran the real test battery in a temporary GitHub Actions workflow on the branch, collected logs, and removed the workflow in the final commit.Clamp verbosity before changing models. Boris Cherny says Opus is strong at long-running work and coding but has a known verbosity quirk; the immediate Claude Code workaround is
claude /config outputStyle=concise.
📡 WHAT SHIPPED
Codex usage reset and efficiency fixes landed. Tibo identified excess usage from images in long sessions with multiple compactions, high p95+ usage from Computer History, and conversation-title generation; the promised full reset covered all paid subscriptions. The follow-up says the reset propagated and fixes for those issues landed.
Kody v2026.08.23 — the homepage now plays an auto-progressing agentic conversation that can be paused, scrolled, and inspected, with tool calls opening inline. The release also makes Kody the stated system of record for memories, email, secrets, packages, and connected surfaces; adds scoped
packages.invokespecifiers; and prevents UI deploys from resetting MCP connections or running jobs.OmaPilot entered alpha. SpencerGBull released the first OmaPilot plugin alpha for OmarchyLinux and is seeking testers; DHH says the prototype runs on a local Qwen 27B model and that a version is planned for Omarchy 4.1. Treat it as a harness experiment, not a production-ready agentic OS.
The surrounding Omarchy ecosystem crossed 1,000 plugins. DHH announced the milestone at omarchyplugins.com, a useful adoption signal for the platform in which OmaPilot is being developed.
Client-artifact warning: @b_nnett reports that Cursor’s Grok bot 0.18.0 shipped with runtime source maps enabled and links reconstructed source and downloads. Audit source-map exposure and other shipped artifacts before putting a cloud-agent client in a sensitive workflow.
🎬 GO DEEPER
🎧 Talking Postgres — 23:01: “Could I explain this to somebody else?”. The episode’s stated production bar is whether you can explain AI-written code to someone else; its show notes frame understanding, rather than code generation, as the bottleneck. Use this as the review-standard segment.
🎧 Talking Postgres — 09:12: red-green TDD makes agents exercise every line. The timestamp is a compact entry point for turning an agent’s implementation loop into an executable test loop.
🧩 Kody v2026.08.23 release notes. Study the concrete boundary work: a visible factory loop, Kody as system of record, scoped package calls, and session isolation across UI deploys.
🧪 smolmachines / smolvm sandbox research. The useful pattern is the escape hatch: detect an environment constraint, move privileged testing to a suitable runner, preserve logs, and clean up the temporary automation.
Editorial take: The winning loop today is not “more autonomy”; it is an explicit handoff, a recoverable environment, and an inspection surface around every consequential run.
Release notes verified
The Kody v2026.08.23 release notes () are available in the bundle and contain concrete changes touching all four requested areas: coding-agent workflows, inspectability, stability, and reusable packages.
Coding-agent workflows
- Kody is the system of record for host agents: Built-in MCP server instructions now tell host agents to prefer Kody memories, email, secrets, packages, and connected surfaces, falling back to a host tool only when Kody lacks the capability, with a reason; assistant state no longer silently forks into Claude memory or Codex notes. ()
- MCP
executerestored: After the platform extraction, MCP execute broke because the platform worker didn't export the fetch gateway; that's restored, and static bearer tokens now reach outbound MCP HTTP requests instead of being silently dropped. () - Scoped
packages.invokespecifiers:packages.invoke('kody:@scope/package/export', options)removes ambiguity when platform and person packages share a Kody id; the bare-ID object overload still works but is marked deprecated; an admin codemod migrates existing source/documentation; privacy-safe telemetry records adoption so the compatibility shim can be retired on evidence. () - Spawned agents default to latest Grok model: Spawned agents and in-repo orchestrators now default to the latest Grok model instead of pinning a version that goes stale. ()
- Fork-outdated merge prompt: When a community listing you forked gets republished, package cards and search results show a yellow 'Fork outdated' button; clicking it copies a prompt your agent can use to pull in upstream changes without discarding your modifications. ()
Inspectability
- Homepage factory-loop conversation player: The 'See the whole loop' link is gone; the homepage plays an auto-progressing agentic conversation you can pause, scroll, and inspect — tool calls open inline, playback resumes from the latest beat, the chat pane reserves its final height, memory results show one-liners instead of raw JSON, the loop ends on a Restart control, mobile taps no longer re-pause after Play, and
prefers-reduced-motionshows the full transcript without autoplay. () - Live fleet code-runs ticker: The homepage shows an above-the-fold counter replaying yesterday's execute totals over the next 24 hours — deterministic, monotonic, and never past the official count; cadence bursts and slows with hashed weights; returning to a background tab snaps to the current number; leftover catch-up stays smooth without tripping the freeze snap. ()
Stability
- UI deploys no longer reset sessions: The origin-facing script no longer owns Durable Objects, so shipping a blog post or Remix change can't drop MCP connections, running jobs, or metering state; the platform split makes this a general guarantee. ()
- Faster/cheaper Vectorize reindexes: Write-time upserts and deploy sweeps skip re-embedding text whose fingerprint hasn't changed; production deploys refresh only builtin capability embeddings instead of rebuilding all user-owned vectors; a
force: truedisaster-recovery path remains after Vectorize data loss. () - e2e wrangler crash fixed: The intermittent test-suite killer is fixed — unread OAuth request bodies are now drained so workerd doesn't terminate the isolate. ()
- Fleet error-rate paging: When the relative error rate on user-package runtime metrics rises sharply, Kody pages admins by email and fans the alert to admin-owned packages, with a six-hour cooldown and no user identifiers, package names, or error strings in the payload. ()
- Contributing docs describe the production worker fleet: The docs now describe the four-script production worker fleet, so agents and humans stop treating a single
npm run deployas a full ship. ()
Reusable packages
- Community package categories: The public community index groups listings under Integrations, Examples, Productivity, Apps, and Utilities; chips filter with
?category=, empty sections hide themselves, and tags remain freeform search keywords. () - Platform-owned listings auto-trust on publish: Official community listings stay trusted through README and icon republishes without an admin re-approving every change; person-owned trust semantics are untouched. ()
- Post-OAuth suggestions match the provider connected: After connecting GitHub, you'll only see packages that actually use GitHub, not a trusted-but-unrelated listing that happens to mention it in a README. ()
- Community OG logos get a white plate: Transparent package marks like AWS and Zendesk no longer disappear into the dark social card; they sit on the same solid white well the listing UI already uses. ()
Gaps, conflicts, and caveats
- The bundle contains only the release notes; no diffs, issue-tracker content beyond PR links, or independent verification were supplied.
- The fleet ticker replays yesterday's totals — it is a simulated replay, not a live event stream ().
-
The
packages.invokebare-ID overload is not removed yet; it is deprecated and shim retirement is gated on telemetry evidence (). - No internal conflicts were found among the release-note claims.
Episode verification
Episode 42 of Talking Postgres, titled "How AI is changing software development with Simon Willison," hosted by Claire, with Simon Willison (creator of Datasette, co-creator of Django). The intro highlights the episode's core themes: the gold standard "Could I explain this to somebody else?" for shipping AI-generated code, engineering management skills for managing AI agents, and the bottleneck shifting from writing code to understanding it, plus the quote "Features are cheap. That doesn't mean you should build them all."
Coding-agent productivity — suggested segments
- [09:12] Red-green TDD makes agents exercise every line
- [18:28] Finding low-hanging fruit among open PRs & issues
- [28:27] Slop proxies add no value at all (likely on evaluating agent output)
- [30:49] Aggressive nitpicking reviews (likely on reviewing AI-generated code)
- [1:00:45] Engineering management skills are so useful (explicitly linked to managing AI agents in the intro)
- [1:18:48] Research agents stopped being absolute garbage
- [1:21:49] The whole point of the "human in the loop"
Cognitive capacity — suggested segments
- [14:42] A million lines mean nothing if you don't understand it
- [23:01] Gold standard: "Could I explain this to somebody else?"
- [48:11] You have to learn to throw things away
- [1:07:53] Writing is thinking, don't outsource it
- [1:08:23] Skill atrophy is a choice you make
Conceptual integrity — suggested segments
- [35:01] Everything in software engineering is about trade-offs
- [51:57] Features are cheap. That doesn't mean you should build them all
- The show notes explicitly link Mythical Man-Month's conceptual integrity concept
Caveat
The bundle contains only the show notes page, not a transcript. Segment titles are the basis for these recommendations; the actual audio must be reviewed to confirm content.
- Simon Willison (firsthand, Talking Postgres podcast): pre-agent engineers wrote ~50–200 lines of debugged production code/day; agents make ~1,000 lines/day possible at the same quality only with senior-level skill, and the new bottleneck is cognitive capacity, so teams are still needed to load-balance it . Cheap agent features erode conceptual integrity — software becomes a Winchester Mystery House of odd additions — so discipline that time-based cost used to enforce must now be deliberate .
- Verification pattern: productively using coding agents means confidently instructing them what to change and confidently verifying the change — not necessarily eyeballing every line .
- Model signal: Simon Willison flags Qwen 3.8 27B as 'a truly astonishing model' after it scored 52 on the Artificial Analysis Intelligence Index — matching GPT-5.6 Luna (max), one point behind GLM-5.2 (max, 753B) and DeepSeek V4 Pro 0813 (max, 1.7T) . Drew Breunig (firsthand): Fable 5 is 'incredible' but pricey; Opus, 5.6, K3, and GLM are 'good enough' for most code, so teams now route work by cost/capability . Ramp's July 2026 Anthropic spend data supports this: Fable 5 = 8.0%, Opus 4.8 = 28.0% .
- Claude Code workflow (Willison, firsthand): exact prompt to Claude Fable 5 in Claude Code for web: 'Put https://smolmachines.com through its paces as a fast secure sandbox. Explore what it would take to use this to run untrusted Python and JavaScript code in a way that is limited in what RAM and CPU time it can take up (protection against "while true") with no network access and filesystem access only to designated files' — goal: execute user-provided tasks for data transformations . The web container lacked /dev/kvm (no nested virt), so the agent ran the real tests on GitHub Actions ubuntu runners (which expose /dev/kvm) via a temporary workflow — an example of Fable being 'relentlessly proactive' .
- Tool release: Bun 1.4 (first stable since the Zig→Rust rewrite) adds Bun.WebView — browser automation via macOS WebKit or Chromium CDP — plus +1,517 Node-compat tests, 2,900+ fixes, 5x lower idle CPU, up to 35% lower memory, and 50% faster Linux startup . Claude Code for web built a shot-scraper-style JSON API on it (load page, execute JS) needing a 192–256MB container for full Chrome on complex pages .
- Linus Torvalds (firsthand Linux commit): in a 'debug session from hell,' the AI repeatedly said the bug was impossible/unsolvable, but kept adding and analyzing debug code when pushed; he let it write the commit message .
- Thomas Ptacek: Stop Making TUIs — agents made native UIs cheap enough that even small personal tools should get real GUIs; Willison confirms from two vibe-coded macOS menu-bar apps he uses daily .
-
Tool releases: llm 0.33 lets
-tbe repeated to combine templates (e.g., savegpt-5.6-lunawithreasoning_effort highaslhigh, thenllm -t lhigh -t pelican) and addsreasoning_summary(auto/concise/detailed) for Responses API models . llm-openrouter 0.7 now uses OpenRouter's Responses API, adds reasoning traces, and adds server-side Shell/WebFetch/WebSearch tools (enable with-T WebSearch) .
@threepointone posted a satirical takedown of 'agent factory' hype, which Theo boosted with a one-word reply: 'Man' . In the skit, a builder boasts an 'incredible factory' that is 'highly optimised, fully automated, zero tolerance for defects and with a continuous feedback cycle', async-interactable from phone/laptop/messenger with shared context that is 'always learning' , and where 'every agent has full context, can spawn other agents, review their work, fix defects, and ship continuously' — but repeatedly cannot say what it builds: only 'Software', 'right now we're mostly using it to improve the factory'; eventually the output is 'Anything!', specifically 'a better factory' . The critique: agent meta-loops (spawning agents, continuous shipping, self-improving shared context) are hollow without a concrete end-user product. Anti-hype commentary, not a firsthand workflow.
Matt Pocock is trying out an
/implement-specskill: a multi-agent implementer that takes a spec and tickets, does codebase research in a subagent, implements all tickets in subagents with maximum concurrency, reviews the final code against the spec, and cleans up all worktrees — aiming to "smash out huge chunks of work autonomously with minimal supervision" . Skill file: github.com/mattpocock/skills/....Kent C. Dodds pushes back: being overly prescriptive about such workflows "is a good way to undo optimizations made by the model and harness" — favoring much looser prompts .
Kent will demo his looser-prompt approach in a free @megadevhq workshop: luma.com/mega-its-time.
Codex rate limits (first-party status update from @thsottiaux, Codex & ChatGPT @OpenAI): a tiger team identified the causes of faster-than-expected usage drain for some paid users and is shipping fixes — (a) inefficiencies when using images in long sessions with multiple compactions, (b) high p95+ usage from the Computer History feature, and (c) a conversation-title-generation feature draining more usage than intended . Along with the fixes, OpenAI will do a full reset of usage for all paid subscriptions , scheduled for ~2pm PST the following day (the author first wrote "14pm", then corrected to "2pm") . A day earlier, Codex acknowledged that some users' cache hit rate was worse this week than the prior stable weeks, which can make usage drain faster since consistent cache hits are an important efficiency component; investigation was ongoing . Practical takeaways for Codex users: cache hits are a key usage-efficiency lever, and long image-heavy sessions with multiple compactions plus the Computer History feature cost extra usage.
Theo (@theo) shared a rough tier list of where he'd put every major model right now, posted as an image (https://pbs.twimg.com/media/HQW1x4HaoAAFnXM.jpg) . In a follow-up, he said he was surprised how much people agreed with the list .
@addyosmani flags a trap in agentic dev workflows: "It's easy to get fascinated by the machinery of doing, optimize the machinery and eventually forget what it was supposed to produce. If your software factory mostly produces a better software factory, then you haven't built a software factory. You've built software whose product is itself" . He amplifies @threepointone's viral satire of the "agent factory" pitch — a system that is "highly optimised, fully automated, zero tolerance for defects" where "every agent has full context, can spawn other agents, review their work, fix defects, and ship continuously" — which, when pressed on what it actually makes, yields only "Software," currently used "to improve the factory," i.e., "a better factory" . A credible counter-signal (commentary on a viral post, not a firsthand workflow): watch for the failure mode where coding agents are used mainly to build more coding-agent tooling rather than user-facing software.
FT reports (shared by Simon Willison) that Anthropic's annualized revenue reached $65bn in July 2026, up from $47bn in May; Anthropic expects Q3 profitability and counts 6,000 customers spending $100k+/year annually, per "people with knowledge of the matter" . OpenAI's annualized revenue jumped 35% in the quarter to date to over $40bn, boosted by the GPT 5.6 launch in July . The article introduces the Ramp AI index, which estimates model adoption from billing data across 70,000 companies using Ramp cards . Ramp's July 2026 breakdown of Anthropic model spend: Opus 4.8 at 28.0%, Sonnet 4.6 at 8.3%, Fable 5 at 8.0%, Opus 4.6 at 6.9%, Sonnet 5 at 3.6%, Opus 5 at 3.5% (released July 24), then older models in the low single digits . Willison notes the data looks reasonable given Opus 5's late launch and supports the view that Fable's cost has made it a less popular model .
Drew Breunig (firsthand, in his post "Fable & The End of the Free Lunch") describes a cost-driven shift in coding-agent workflow: before Fable, improving your coding harness or context strategies felt like wasted effort because a new model would arrive at the same price and paper over most problems; Fable was "incredible" but so expensive that Opus (and 5.6, K3, even GLM) was "good enough" for most of the code they needed — so they started deliberately deciding what work went where . His post: https://www.dbreunig.com/2026/08/23/fable-the-end-of-moore-s-law.html
- Riley Brown's Cursor Delegate prompt pattern: a Grokbot agent acts as the mailbox while Cursor's cloud agent does the thinking/writing; when users say "Send this to Cursor," "Ask 5.6 Sol," or "Delegate this," the delegate returns the cloud agent's full message back verbatim, not a paraphrase .
- Setup detail: delegated agents get a private scratch repository for writing and replying — it is not a product codebase unless the prompt explicitly asks the agent to build something .
- Model config: default model is GPT-5.6 Sol (gpt-5.6-sol); if the user requests "high" or "5.6 Sol high," set model_params.reasoning to high, otherwise omit extra model parameters unless the user names them .
Kent C. Dodds' open-source coding agent Kody released v2026.08.23 . New in this release: the homepage plays an auto-progressing "factory-loop" agentic conversation you can pause, scroll, and inspect, with tool calls opening inline; the "See the whole loop" link was removed . Dodds says this release makes Kody "much more stable," brings "tons of new official packages to fork so you don't have to build things from scratch," and moves the project "much closer to release" . Firsthand update from the project creator.
@b_nnett discovered that the Cursor team shipped Grok bot 0.18.0 with runtime source maps enabled, allowing source code reconstruction; they posted the reconstructed source at github.com/b-nnett/grok-bot-0.18-reconstructed. Theo (t3.gg) adds that the same team also accidentally left the entire T3 Code GitHub repo public .
Boris Cherny (@bcherny), speaking for the Claude team, says Opus excels at long-running work and coding but has a known verbosity quirk that the team is prioritizing to fix . He shipped a quick band-aid in Claude Code: run claude /config outputStyle=concise to reduce verbosity immediately, without waiting for the longer-term fix . He notes every model has different strengths and quirks .
OpenAI Codex (via @thsottiaux) shipped usage/rate-limit fixes: all paid subscriptions received a full usage reset; fixes landed for usage inefficiencies from images in long sessions with multiple compactions, high p95+ usage for Computer History, and a conversation-title generation feature draining usage; more fixes coming, plus a novel efficiency approach to work on next week .
@rileybrown (practitioner building agents) says he uses Codex with "5.6 sol" for almost all document creation and app building, repeatedly tries to migrate to GPT Work but always ends up remoting into desktop Codex from his laptop. He wants no separation between coding and knowledge-work tools, and wishes there were a cloud Codex identical to the desktop version . He calls Codex "the best AI product ever released" .
SpencerGBull released the first alpha of OmaPilot, a plugin for OmarchyLinux that makes it an "Agentic Ready OS," and is seeking testers . DHH says the demo runs off a local Qwen 27B model and a version will ship in Omarchy 4.1 .
Codex app tip: use side chat to keep working while the main agent thread is still running — @derrickcchoi says it's "really underrated when you're waiting on the main thread to finish cooking" , and @rileybrown says he side-chats "more than main chat" . Both are firsthand practitioner reports.
Riley Brown (@rileybrown) argues the three winning AI interaction paradigms will be chat (ChatGPT, Claude), bots/agents (Grokbot), and code (Codex, Claude Code). He considers Claude Cowork and GPT Work too similar to chat to be distinct paradigms, predicts every lab should copy the GrokBot UI, and expects a fourth 'team agent' paradigm to emerge, with Claude Tag as the closest current example .
DHH announced "ONE THOUSAND OMARCHY PLUGINS! 🍾🎆" — Omarchy has reached 1,000 plugins, linking to https://omarchyplugins.com.
Conceptual integrity and counting lines of code
In this newsletter:
- Conceptual integrity and counting lines of code
Plus 6 links and 4 quotations and 1 note and 3 releases and 2 research reports
Sponsor message: Constrain AI agent abilities with trusted runtimes and Agentic Classifiers
AI agents run 24/7 and will try anything to achieve their goal – at 100x the scale of a human workforce. Teleport (opens in new tab) helps you deploy AI safely: starting with isolated ephemeral trusted runtimes, followed by a deep AI audit. Agentic Classifiers take it from there, re-reviewing sessions to check and reclassify behavior using downstream meta-audit events that can lock or escalate AI agents’ actions.
Last week I recorded an episode of the Talking Postgres podcast (opens in new tab) with Claire Giordano on the subject of “How AI is changing software development”. We had a really great conversation. Here are a couple of my highlights from a lightly edited transcript (prompt to Claude: “very minor edits to remove disfluencies”).
This is the latest version of an argument I’ve been trying to build about why sometimes it does make sense to talk about lines of code as an indicator of productivity with coding agents, at 35:01 (opens in new tab):
A lot of people will tell you it makes no sense to measure productivity in lines of code. I’d actually disagree, because there’s a hard limit. In the before-times, a software engineer could produce a few hundred lines of production-ready code per day — and 200 lines of working, debugged, production-level code is an incredibly good day. Most days you’d produce 50 or 60.
If agents let you produce a thousand lines of debugged code, that really is a very meaningful improvement — as long as the code is the same quality: maintainable, tested, all of that. You can get to that point with agents, but it takes a huge amount of skill and knowledge and experience. That’s what senior engineers are made of.
I can do way more work as a single engineer than I could without agents. So you could argue, why should a company have more than one engineer? Beyond the obvious bus factor thing — a team of one is a very badly designed team — the answer is that the new limiting factor is cognitive capacity. I can churn out code a hundred times faster. I don’t have the cognitive capacity to stay on top of 100 times the amount of code. So you still need a team of engineers, so you can load balance that cognitive capacity across the team.
And this section on conceptual integrity at 46:03 (opens in new tab), which Claire equated to the Winchester Mystery House (opens in new tab)!
Simon: There’s a concept in The Mythical Man-Month — conceptual integrity — where well-designed software has an integrity to it: there are no surprises in it, it covers exactly the right domain of things, everything fits together and makes sense. That’s so much harder with coding agents, where you can have an idea for a feature, run a prompt, and five minuteslater you’ve got the feature. Your software grows little weird bumps in funny different directions.
Claire: You know my analogy for that? The Winchester Mystery House.
Simon: It’s got 140 rooms, because the woman who built it was the widow of the guy who invented the Winchester rifle, and her psychic told her she’d be haunted by the ghosts of everyone killed with that rifle unless she kept building the house forever. So for 40 years she kept adding new rooms. That’s exactly the problem with coding agents and software: it’s very easy to keep adding new rooms, because the cost of adding those rooms is so much cheaper. What you end up with is something where the conceptual integrity falls apart — and then it’s harder to make decisions about it.
It all keeps coming back to discipline. It used to be that the discipline was enforced on you by the amount of time it took. You’d come up with an idea for a crazy feature and think “yeah, but that would take me a week — I cannot justify that, so I’ll forget about it.” If it takes an hour, it’s so much easier to justify.
(Side-note: the Wikipedia article includes credible sources that dispute the story about the psychic.)
Link 2026-08-17 We Tracked a Shipment of Rare Books. It Ended at an Amazon AI Training Facility (opens in new tab):
Excellent piece of reporting from 404 Media. For a while now there have been stories of book dealers receiving orders for large volumes of books from apparently price-insensitive anonymous customers, widely suspected to be companies looking to scan them for AI training (see my previous coverage (opens in new tab)of Anthropic’s book scanning from June 2025.)
404 Media investigated with an AirTag!
In July, one bookseller told me they received a very large order of around 1,000 books on Biblio, one of these marketplaces. The seller agreed to put an Apple AirTag provided by 404 Media in one of the books included in this order so we could see where the book was going. And by extension, which company, AI or otherwise, was behind this massive order.
The book ended up delivered to the VGT3 corner of the LAS8 Amazon facility (opens in new tab) in the north east of Las Vegas, where the entrance carried this on-the-nose logo of a dinosaur with a book!

Photo credit: 404 Media
Online forum discussions between Amazon workers confirmed that VGT3 destructively scans large volumes of books.
Link 2026-08-17 Qwen 3.8 27B scores 52 on the Artificial Analysis Intelligence Index (opens in new tab):
That’s the same score as GPT-5.6 Luna (max), and just one point behind GLM-5.2 (max) and DeepSeek V4 Pro 0813 (max) - that GLM is 753B (opens in new tab) and that DeepSeek is 1.7T parameters (opens in new tab), and Luna is size unknown but presumably a whole lot bigger than 27B.
Qwen 3.8 27B is a truly astonishing model (opens in new tab).
Link 2026-08-18 Mojo🔥 is now open source (opens in new tab):
The Mojo programming language has been promising an open source release since May 2023 (opens in new tab). Last week they shipped their 1.0 (opens in new tab) and today they have followed through on that original promise, releasing the compiler and toolchain under an Apache 2 license.
When Mojo first launched the stated goal was to produce a superset of Python, so existing Python code could be used to bootstrap their own ecosystem. That plan changed around August 2025 (opens in new tab):
Mojo may or may not evolve into a full superset of Python, and it’s okay if it doesn’t.
We’re encouraged by how well AI-assisted coding tools already help migrate Python to Mojo today, and we’re confident that future tooling and ecosystem maturity will make this evolution even smoother.
Today Mojo is its own language, optimized to make GPU programming as painless as possible using syntax inspired by Python, if not 100% compatible with existing code.
Quote 2026-08-19
My hypothesis is that there is a new opportunity for Extensible Software on the web. LLMs radically lower the cost of authoring extensions, and modern sandbox primitives lower the deployment cost and provide good security boundaries. We can build our app as a solid, accountable core, and allow users to safely extend it in many directions by having LLMs fill in the missing pieces. We can give our users super powers.
Jeremy Morrell (opens in new tab), Extensible Software in the age of LLMs
I tasked Claude Fable 5 running in Claude Code for web with the following research task:
Put https://smolmachines.com through its paces as a fast secure sandbox. Explore what it would take to use this to run untrusted Python and JavaScript code in a way that is limited in what RAM and CPU time it can take up (protection against "while true") with no network access and filesystem access only to designated files
Goal is to be able to use this to execute user-provided tasks for things like data transformations
It quickly ran into a problem: the Claude Code for web environment can’t run smol machines (opens in new tab). Quoting the notes it wrote (opens in new tab):
This Claude Code container: Linux 6.18.5-fc-v20 (itself a Firecracker guest), 4 vCPU, 15GB RAM. No /dev/kvm, no vmx/svm CPU flags → no nested virt.
smolvm machine run fails as expected: “kvm not available”.
Plan B: GitHub Actions ubuntu runners DO expose /dev/kvm → run the real test battery via a temporary workflow on this branch, collect logs, remove workflow in final commit.
And Plan B is what it did (opens in new tab), installing smolvm and running these tests (opens in new tab) directly in a GitHub Actions runner against that branch.
That was a creative solution to the environmental limits posed by Claude Code for web. Another example of Fable being relentlessly proactive (opens in new tab).
Today saw the long awaited release of Bun 1.4 (opens in new tab), the first stable version since the infamous Rust rewrite a few months ago (opens in new tab).
Interestingly, the Rust rewrite was downplayed in the release notes, which introduced a bewildering array of new features and claimed 2,900 additional bug fixes:
Bun 1.4 adds +1,517 tests from the Node.js test suite - our biggest jump in Node.js compatibility since Bun 1.0. Bun v1.4 also fixes over 2,900 issues. It reduces idle CPU usage by 5x, reduces memory usage by up to 35%, and starts 50% faster on Linux. It adds
Bun.Image,Bun.WebView,Bun.markdown,Bun.cron(),Bun.Terminal,bun run --parallel,bun test --parallel,bun audit fix,bun dedupe, andbun prune. And it rewrites Bun from Zig to Rust.
Of these the one that most caught my eye was Bun.WebView, which adds first class support for browser automation to Bun core using either macOS WebKit or control of a local Chromium process via the Chrome DevTools Protocol (CDP).
I had Claude Code for web build a prototype of a web API providing the ability to load a web page and then execute JavaScript against it, inspired by my shot-scraper javascript (opens in new tab) CLI tool - partly to see how much RAM would be needed by such a service.
Here’s that TypeScript server implementation (opens in new tab), which appears to need a 192MB-256MB container to run a full Chrome against complex web pages - tested using cgroups.
Promptwatch is part of the emerging “GEO” space, for Generative Engine Optimization - the chatbot version of SEO, where companies offer tools and consulting to help your site increase its presence in replies to prompts inside tools like ChatGPT.
The Promptwatch product uses automation to track responses to prompts across end-user chat products like ChatGPT, Claude, and Gemini. They publish aggregate reports on this as part of their own content marketing strategy, which do seem to provide credible hints as to otherwise invisible design changes to those products.
Their own tracking shows a notable change aligned with the GPT-5.6 rollout earlier this month:
The percentage of all ChatGPT Search fanout queries that contain the site:operator, per day. The share hovered between 0.3% and 0.5% for weeks, dipped briefly to 0.15% on August 3 to 5 (consistent with a staged rollout or pre-launch experiment), then jumped to 16-17% on August 8.
It’s important to note that these figures only reflect the prompts for which they have automated tracking enabled.
This corresponds to OpenAI’s somewhat vague August 6th announcement (opens in new tab):
For Plus and Pro users, we’re updating GPT‑5.6 Sol in Chat to be more reliable with facts and provide more focused answers.
Once again I am hampered by OpenAI’s decision to actively obscure their system prompts, but from poking at ChatGPT I believe their latest search tool has a shape like search(query, recency, domains) rather than encouraging a site: operator directly.
In a follow-up (opens in new tab) on August 18th Promptwatch reported that ChatGPT appeared to have greatly reduced the likelihood of Reddit being used in those searches. My own attempts to ascertain if the system prompt has been updated to discourage Reddit sourcing have been unsuccessful - the most thorough leaked system prompt (opens in new tab) collection I know of doesn’t yet show any relevant changes.
Quote 2026-08-21
After I released version 1.0, I figured I would have to do the rotations myself. So I sat down with ChatGPT and I didn’t get it to write the code, but I got it to educate me. With a patient, interactive tutor, I was able to finally do what I hadn’t by reading books and asking mathematician friends – I learnt how to use quaternions just enough to make the app work.
So learning doesn’t stop just because I outsource a bunch of thinking to AI. It pushes me to learn more. I like that as an outcome.
Matt Webb (opens in new tab), Galactic Compass 2: now with new augmented reality mode
Link 2026-08-21 Stop Making TUIs (opens in new tab):
Thomas Ptacek advocates for building real native user interfaces for even the smallest of personal tools, because coding agents have reduced the cost of getting a usable-enough GUI up and running to almost nothing.
I wrote about my vibe-coded bandwidth and GPU monitoring macOS task bar apps back in March (opens in new tab), and I’m still using both of those on a daily basis.
I’m not habitually knocking out real UIs for my other projects yet, but I’m running out of excuses!
Thomas:
If you haven’t tried your hand at turning one of your 500 throwaway CLIs into a native app, you’re doing yourself a disservice. Go build a native UI. It’ll probably change the way you think.
Now that this plugin is compatible with LLM 0.32 (opens in new tab) it can display the reasoning traces for LLMs available through OpenRouter.
Updated for compatibility with LLM 0.32 (opens in new tab).
Models now use OpenRouter’s implementation of the Responses API (opens in new tab).
Three new server-side tools: Shell (opens in new tab), WebFetch (opens in new tab), and WebSearch (opens in new tab). Enable these with options like -T WebSearch.
Release: llm 0.32.1 (opens in new tab)
Fresh installs of LLM stopped working the other day because the OpenAI Python library dropped its usage of httpx, and it turned out LLM depended on that library but only installed it via a transitive openai dependency.
This dot-release fixes that for the moment by pinning to openai<3, and a soon-to-drop 0.33 release will switch from httpx to httpx2 (opens in new tab).
The key skill required to make productive use of coding agents is being able to confidently instruct them on how to make changes and then confidently verify that those changes have been applied in the correct way.
Sometimes this involves reviewing every line of code they have written, but there are other ways to achieve that goal. Eyeballing every line of code has never been the most effective way to validate a chance to a piece of software.
Release: llm 0.33 (opens in new tab)
My highlights from this release:
- Upgraded to the OpenAI Python library 3.x and switched the HTTP client dependency from
httpxtohttpx2. #1608 (opens in new tab), #1631 (opens in new tab)
I shipped a quick 0.32.1 fix (opens in new tab) for this yesterday, but this is the more comprehensive fix.
llm embedandllm embed-multinow accept--key. The PythonEmbeddingModel.embed(),EmbeddingModel.embed_multi(),Collection.embed()andCollection.embed_multi()methods acceptkey=too, passing the resolved per-call key to embedding plugins without changing shared model state. Existing plugins that readself.keycontinue to work through a compatibility fallback. Thanks, ChrisJr404 (opens in new tab). #757 (opens in new tab), #1620 (opens in new tab)
The embedding models now use the same pattern for keys that regular LLM models do.
llm prompt -t/--templatecan now be repeated to combine templates in order. This allows model configuration and options from one template to be used with a prompt from another.
This unlocks a neat pattern where you can create templates that package a model with a set of default options:
llm -m gpt-5.6-luna -o reasoning_effort high --save lhigh
llm "Generate an SVG of a pelican riding a bicycle" --save pelican
# Combine and run the templates
llm -t lhigh -t pelican
- Reasoning-capable Responses API models now support a
reasoning_summaryoption withauto,concise, anddetailedvalues. This can be used with llm openai endpoint –responses (opens in new tab). #1600 (opens in new tab)
This is particularly useful for exercising different models that provide their own imitation of the OpenAI Responses API.
Quote 2026-08-22
And this was a debug session from hell, enormously helped by an AI doing much of the grunt-work.
I’d like to call it my tireless helper, but the AI several times stated flat out that this was impossible and unsolvable and that we should just write a report about it.
I suspect those things have been trained by people who may not be quite as stubborn as I am.
But while the AI was ready to give up several times, it did keep adding debug code and analyzing it faithfully when I pushed. So credit where credit is due and I let the AI write the commit message above.
Linus Torvalds (opens in new tab), drm/xe: Don’t hand out the flat CCS storage as usable VRAM
Quote 2026-08-23
Prior to Fable, it felt silly to waste too much time improving your coding harness or context strategies. A new model would arrive at the same price (or cheaper!) and paper over most of your problems.
But then Fable landed. It was (and still is!) incredible. But the cost was so high and Opus was good enough (as was 5.6, K3, and even GLM) for most of the code we needed.
So we started to think about what work went where.
Drew Breunig (opens in new tab), Fable & The End of the Free Lunch
Link 2026-08-23 Anthropic’s best AI model struggles to attract users as cheaper tools thrive (opens in new tab):
A few interesting numbers in this FT story gathered from “people with knowledge of the matter”:
Anthropic’s “annualized revenue” for July is up to \$65bn - it was \$47bn in May, and I collected more historic numbers here (opens in new tab).
Anthropic expect Q3 to be profitable according to the same model they used to declare Q2 profitable. “It also told investors that it had 6,000 customers that spend \$100,000 annually or more.”
As for OpenAI, “annualised revenue has jumped 35 per cent in the quarter to date and is now over \$40bn, with the launch of GPT 5.6 in July jolting the company’s performance after a sluggish start to the year”.
This article also introduced me to the Ramp AI index (opens in new tab), which uses billing data from 70,000 Ramp credit card using companies to estimate model adoption.
Here’s Ramp’s breakdown of Anthropic model spend for July 2026, which looks reasonable given that Opus 5 was only released on July 24th, and supports the idea that Fable’s cost has made it a less popular model:
Opus 4.8: 28.0%
Sonnet 4.6: 8.3%
Fable 5: 8.0%
Opus 4.6: 6.9%
Sonnet 5: 3.6%
Opus 5: 3.5%
Opus 4.7: 1.7%
Sonnet 4.5: 1.3%
Haiku 4.5: 1.0%
Opus 4.5: 0.7%
If you find this newsletter useful, please consider sponsoring me via GitHub (opens in new tab). \$10/month and higher sponsors get a monthly newsletter with my summary of the most important trends of the past 30 days - here are previews from April (opens in new tab) and May (opens in new tab) and June (opens in new tab).
- Simon Willison (firsthand, Talking Postgres podcast): pre-agent engineers wrote ~50–200 lines of debugged production code/day; agents make ~1,000 lines/day possible at the same quality only with senior-level skill, and the new bottleneck is cognitive capacity, so teams are still needed to load-balance it . Cheap agent features erode conceptual integrity — software becomes a Winchester Mystery House of odd additions — so discipline that time-based cost used to enforce must now be deliberate .
- Verification pattern: productively using coding agents means confidently instructing them what to change and confidently verifying the change — not necessarily eyeballing every line .
- Model signal: Simon Willison flags Qwen 3.8 27B as 'a truly astonishing model' after it scored 52 on the Artificial Analysis Intelligence Index — matching GPT-5.6 Luna (max), one point behind GLM-5.2 (max, 753B) and DeepSeek V4 Pro 0813 (max, 1.7T) . Drew Breunig (firsthand): Fable 5 is 'incredible' but pricey; Opus, 5.6, K3, and GLM are 'good enough' for most code, so teams now route work by cost/capability . Ramp's July 2026 Anthropic spend data supports this: Fable 5 = 8.0%, Opus 4.8 = 28.0% .
- Claude Code workflow (Willison, firsthand): exact prompt to Claude Fable 5 in Claude Code for web: 'Put https://smolmachines.com through its paces as a fast secure sandbox. Explore what it would take to use this to run untrusted Python and JavaScript code in a way that is limited in what RAM and CPU time it can take up (protection against "while true") with no network access and filesystem access only to designated files' — goal: execute user-provided tasks for data transformations . The web container lacked /dev/kvm (no nested virt), so the agent ran the real tests on GitHub Actions ubuntu runners (which expose /dev/kvm) via a temporary workflow — an example of Fable being 'relentlessly proactive' .
- Tool release: Bun 1.4 (first stable since the Zig→Rust rewrite) adds Bun.WebView — browser automation via macOS WebKit or Chromium CDP — plus +1,517 Node-compat tests, 2,900+ fixes, 5x lower idle CPU, up to 35% lower memory, and 50% faster Linux startup . Claude Code for web built a shot-scraper-style JSON API on it (load page, execute JS) needing a 192–256MB container for full Chrome on complex pages .
- Linus Torvalds (firsthand Linux commit): in a 'debug session from hell,' the AI repeatedly said the bug was impossible/unsolvable, but kept adding and analyzing debug code when pushed; he let it write the commit message .
- Thomas Ptacek: Stop Making TUIs — agents made native UIs cheap enough that even small personal tools should get real GUIs; Willison confirms from two vibe-coded macOS menu-bar apps he uses daily .
-
Tool releases: llm 0.33 lets
-tbe repeated to combine templates (e.g., savegpt-5.6-lunawithreasoning_effort highaslhigh, thenllm -t lhigh -t pelican) and addsreasoning_summary(auto/concise/detailed) for Responses API models . llm-openrouter 0.7 now uses OpenRouter's Responses API, adds reasoning traces, and adds server-side Shell/WebFetch/WebSearch tools (enable with-T WebSearch) .