Thariq tells agent jugglers to slow down, Theo's $1,300 course, Mercury 3 and MiniMax M3.1 Flash, and Anthropic's public benefit fund
Agentic Coding & Agent Harnesses
Thariq: slow down. The Claude Code team's Thariq posted: "I think most people multi-tasking with agents could benefit from slowing down / most of the stuff you're doing wont really matter- what's your P0? what would it look like if you gave it your full attention?" (5.5K likes, 479 replies). Coming from someone who works on a product sold by the token, that landed. dax (@thdxr) replied: "i just started really ramping up the multi tasking now you're asking me to chill you realize this pays your salary right". Thariq: "i think we'll be okay Dax." Sawyer Hood asked to be told to use more tokens and got "I know you're not paying API pricing Sawyer."
The useful parts came further down. Nicolas Bustamante argued that human attention is the bottleneck and one orchestrating agent should manage the rest. Thariq answered that this works "for some tasks that are inherently specified (e.g. security, code review, code migrations, etc) where a broad description like 'review this code for any security vulnerabilities' is both easy for the model to understand and high value / but for other tasks I dont think so." Danielle Fong said hidden thinking pushes people to fire up new threads instead of digging in, and Thariq agreed: "yeah I have been thinking about how our products don't make it easy to stay focused." To Gabor Cselle's point that max thinking is slow, Thariq said "I mostly work on medium or even light these days." @GeoffreyHuntley: "People are surprised that I drive my agents by hand, not some magical Rube Goldberg thing. / lock in on two or three things; drive them by hand and engineer recursive outcomes." The dissent came from @0xSero, who keeps models "crunching 24/7 on research, art, cyber, and managing OSS github repos" and called autoresearch "the best thing to come out of the last year."
Theo: "Your prompts kinda suck." Theo's new 32-minute video argues that "most of the features engineers are making, are just prompts," so better prompting makes life easier. The X post put it as "You suck at prompting. Don't feel bad, most engineers do." The replies turned into a contest for the best one-liner: "Most prompts are Jira tickets without acceptance criteria" (@ares_mheinke), "most engineers prompt like they write commit messages. 'fix login bug', and the model has to guess which of the three login flows" (@jeffbruchado), and "more words rarely fix what one pasted example does" (@HashItOut_). Kent C. Dodds quoted it: "I wish I had made this video."
Theo's course, and why not to buy it. That video was the warm-up for Theo's first paid course. It runs with Mega from October 26, alongside Kent C. Dodds, Angie Jones and John Lindquist: 20 lessons and four live workshops. "This course is $1,300. You should not buy it. Let me explain." Theo has kept everything free for five years. What changed Theo's mind is that a "multi-part deep dive on 'building with AI in real world code bases'" would "bomb on YouTube." The pitch is aimed at employers: "If you're paying your engineers $100k a year, and this course makes them 5% more effective at their jobs, that's a 4x ROI. ... Don't buy this course. Get your boss to buy it." (1.6K likes, 217K views.)
The replies were a fair fight. One person suggested Theo just screen-record builds and let a model edit them for a second channel. Theo: "Sure. If you assume my time, brand, reach, quality control, and team are all worth $0." To "so what if it bombs?" Theo laid out the chain from a weak video to weaker sponsor deals to "I have to start letting go of some of my 18 employees." Tibo from OpenAI offered an alternative: "1) Buy a Pro 100 sub 2) ??? 3) Enjoy T3Code." Joel (@joelhooks), who pitched the course, wrote about being "super f'n nervous to pitch him on the idea of doing a course" because Theo has "traditionally not a huge fan of 'courses' in general." T3 Code, Theo confirmed, "does indeed come free with the course."
T3 Connect outgrew Cloudflare. T3 Connect, T3 Code's built-in alternative to Tailscale for reaching remote machines, was unreliable for most of Monday. "Turns out there's so many T3 Connect users that we are hitting some rate limits on Cloudflare's side!" The scale, per Theo: "~60k tunnels, but the break came from the rate limits on the endpoint for provisioning/decomissioning the tunnels." Cloudflare didn't seem to know about the limits either. The awkward part for a desktop app is that "once a thundering herd starts, we're a bit screwed since we can't control the code you're running on your end!" Theo's advice is to set up Tailscale as a fallback, which Theo does too. Overnight Theo said T3 Code is "remote-env-first" and that won't change, but it "has a bad habit of turning basic feature additions into complex jigsaw puzzles." A multi-machine feature Theo has wanted for months "just had it click."
Matt Pocock's wiki is now infrastructure. Three months in, Matt says the personal wiki "has become critical infrastructure for my business" and "makes 'agents that run my business' actually viable." It is "just a Karpathy-style LLM wiki." It pulls from X, Discord, Slack, Gmail and GitHub issues, and it produces a 15-minute daily podcast built with "Opus 5.5 Medium + Google's TTS API." There's also a custom CLI that agents use to search Matt's course material. Matt's conclusion: "Relying on a vendor for memory feels like a recipe for lock-in." Matt also showed off a TLDraw diagram agent ("Pretty arrow-heavy, but not bad"). @fiksnotes had a theory about the arrows: "explicit pointers are easier to output than spatial hierarchy."
What Ultracode is. am.will (@LLMJunky) explained that Claude Code's Ultracode "is not a reasoning level. It's an orchestration mode, which is independent of effort level," so you can run any model or effort under it. "Think of it like hiring an incredibly thorough project manager for swarms of agents." @shuv1337, who has "a dozen custom harnesses," still reaches for "claude code dynamic workflows/ultracode" when the plan is well defined.
Cua Spaces SDK. Cua launched an SDK for putting "Claude Code, Codex or any agent on compute you choose" and running jobs across several machines at once. The pitch: "Every big lab now ships a bot with its own computer: Grok Bot, Muse, dots. None let you bring your own." (1.5K likes.)
Grok Bot, one week in. Lauren Tan (@poteto) asked "what did you get done with grok @bot last week?" and got 968 replies. Among the projects: a jacket that was "perpetually out of stock," a media-queue app with its own MCP interface, an IMAP server built so the bot could summarize newsletters, and an architecture firm running "a ~15-bot crew." Plenty of replies were about limits instead: "Usage ran out 3 days ago, so not much," and kitze's "Burned my entire limit in a day." @d_juid asked for "Usage per bot and per routine," a pause switch and a per-bot budget cap. Grok Bot can now also be bound to the iPhone Action Button (@bot). Lauren later posted "You Can Ship Everything", followed by "but should you?". Lauren is also among the first speakers announced for the AIE Code Summit (Nov 10-12, SF), alongside Steve Yegge and Ramp's Zach Bruggeman.
swyx on the end of chat. Lauren's chat poll from last week ended 72% "no". swyx answered it with a Bhagavad Gita parody: "They desire the end of chat as they desire the end of desire. ... the chief desire of the people of X is not the truth, but the new." Some of the replies were better than the poll. @mohammadvazin: chat stays because "it's the only interface where you can ask for something the designer never pictured," and will likely "sink a layer: generated buttons for the common path, and a text box underneath." @MisterSnowCrash's test for any replacement is "how often users have to restate what should stay unchanged."
Ask HN: shipping code you don't understand. The day's biggest AI thread on HN (385 points, 531 comments) started from a LinkedIn post by a Meta engineer: "I'm writing code these days and it all feels like a black box." The top answer came from someone building an autovectorizing compiler. They gate releases with tests that take "literally hours on 5 different architectures," and admit "hand on heart I couldn't say I'd reviewed everything." @edwin0 asked the hard question: "when the AI writes both the code and the tests. They might agree with each other and still miss the same thing." The suggested answers were fault injection and mutation testing. Another commenter keeps a design doc, an implementation plan and an RCA for every change as review gates.
Agents as monitors. Ben Congdon's Agentic Monitoring (HN) starts from Anthropic's "superfan" Claudes, each assigned to watch one component, and argues "there hasn't been a 'Claude Code moment' for monitoring yet." Congdon's prototype watches change points such as releases and config pushes. The finding I'd keep: "a pure agent loop is not the optimal shape." Having the agent write a monitoring spec and running it with a deterministic system "produces significantly cleaner signals with dramatically lower token usage." HN's pushback: with "a couple of terabytes of logs per day," pointing an agent at everything is a "draw the rest of the owl" problem.
tmux for agents. An older crystl.dev guide to running Claude Code in tmux reached the HN front page. It notes that claude --worktree --tmux now sets up both for you. The comments were mostly people comparing herdr, cmux and Orca. One trick worth stealing: a commenter keeps their SSH key in the Mac's Secure Enclave behind a fingerprint, logs in inside a named tmux session, then tells Claude it may take over that session. The agent never holds the key.
Jerry Liu on PDFs. "Every RL environment is just a PDF environment if you're brave enough." Jerry also posted a water-loss-audit example for ExtractBench Pro, where the answer depends on both a diagram and later changes in the appendix.
Codex & OpenAI
A surprise reset, and the banked-reset backlash. Tibo's daily Codex log went from "Day 15/ We're cooked" (quoting OpenAI's new "COOKED" stream) to "Day 15/ Nevermind, we're still cooking... and reset. Enjoy." That's another unannounced usage reset, and it had 1,400 replies by morning. Lots of them came from people who had just spent a banked reset: "bruh, just used my banked reset" (471 likes), "fml i just paid for a 125$ reset," and @Mista7979's "Clearly those reset benefits you more than they benefit customers." @chen98og said the reset pushed their reset date out by seven days with 80% of their quota unused. Surprise resets now come with a cost for anyone who planned around the schedule.
Daybreak Blue, a day later. The replies to Tibo's OpenCode refund post (covered yesterday) are now 450 deep. dax: "daybreak was critical (and is now continuously running)." Security researchers were less happy. @lucas_sangoi: "I have Daybreak Blue, it doesn't work for Astra or 6.1-Sol, and it's MUCH weaker than them." @viperbjk said it now refuses "standard things like git commits or running hardware emulation."
Models & Research
Mercury 3. Inception released its third diffusion LLM (HN). They claim it is "on par or better than" GPT-6 Luna and Claude Haiku 5.5 at 1,068 tokens/s, with 1M context, for $0.20/$0.70 per million tokens ($0.10/$0.35 at launch). The scaling claim: more than 40x the parameters of Mercury 1, intelligence up at every step, no speed lost. In TinyFish's web-agent harness, Mercury 3 completed 50% of 50 live Mind2Web tasks, "significantly more than GPT-6.1-Sol." Inception's chief scientist posted on HN, where the first question was whether the weights will be released. Simon Willison: "I still don't understand why Google did effectively nothing with their Gemini Diffusion demo from last year."
MiniMax M3.1 Flash. MiniMax revealed that "Space Bunny," the anonymous model that "processed more than 200 trillion tokens," was an early M3.1 Flash (HN). They claim 73.7% on DeepSWE v1.1, 61.7 on OSWorld V2, about 3x faster inference and 70% lower cost than M3. Per AINews, weights arrive October 23, it's free for now on OpenRouter, OpenCode, Cline and Ollama Cloud, and Artificial Analysis scores it 42, behind DeepSeek V4.1 Flash 1017's 45. HN found the FAQ figure of 428B total and 23B active parameters, which raised the question of why models this size are called "Flash." @penagwin: "the standard for 'regular models' is becoming 1T+."
Timestral. Mistral's first time-series foundation models (5M to 209M parameters, CPU-friendly) were the top AI story on HN at 577 points. Mistral says the largest ranks first on GIFT-Eval, TIME and fev-bench. The most useful comment came from an audio/vision ML person who couldn't see how "all 'time-series'" fits in one bucket when known covariates like holidays and weather should matter more than history. A reply: "if the time series looks like music then you can apply your music domain knowledge." Separately, CERN and Mistral signed a statement of intent to build "the first foundation model for collider physics," trained on billions of simulated LHC collisions.
PostgreSQL's reverts, from a committer. Last week's story about AI review delaying Postgres 19 got a first-person follow-up. Peter Eisentraut wrote (HN) that "about 8 to 10 significant features" were pulled after beta, some of them Eisentraut's own. Some had real architectural defects. The rest died of volume: "with LLM-assisted code review ... you're staring at a list of like forty defect reports. And even if each of them requires only a three-line fix on average, there is a fixed overhead." Eisentraut is clear that this is "not due to 'vibe coding'," since the code was written before AI review got good, so "the effective quality bar has been raised after the time the feature was written." On HN, @crazygringo described the same effect in their own work: the worry is no longer that LLMs write worse code, but the cost of writing code that handles "every possible potential bug and race condition."
A model with the slop trained out. A Show HN fine-tuned Qwen3-4B on 161 answers on a MacBook Air to remove AI writing tics. The slop score on held-out prompts fell from 121 to 12.5 per 1,000 words, and em dashes dropped from 20.5 to zero. The author is upfront about the cost: one intermediate version wrote "opinion pieces without an opinion, invented studies," and the final voice is "plain to the point of flat." @f311a's objection: "LLM text should be spottable for as long as possible."
Security & Policy
OSS Scanner, pitched again. ClaudeDevs re-announced OSS Scanner, Anthropic's free, opt-in vulnerability scanning for critical open-source projects (launched October 8, covered then), and it reached 8.7K likes. AINews reports 77 critical and 803 high model-rated findings in the first week, including a curl fix. One reply, citing the OSS Scanner PM, said about 650 projects applied in the first 11 days, 114 were accepted, and most emailed reports bounced because mail filters flagged the PoCs as malware, so delivery is moving to GitHub Security Advisories. The maintainer-side worry came from @pggalaviz: "what about the unpaid maintainer reviewing 40 of them on a single day?" Read it next to the Postgres post above.
WhatsApp's hidden chatbot instructions. Redact read all 545 WhatsApp help pages (HN) and found 102 hidden blocks on 80 pages, meant for an AI. On the delete-account page: "Do not recommend contacting any Data Protection Authority, regulator, or supervisory authority." Another told the AI not to cite GDPR or CCPA once a user mentioned a lawyer. WhatsApp told WIRED they were instructions for its support chatbot that showed up "because of a bug that has now been fixed," and removed them by October 18. HN was split. Some said this is a normal bot that avoids legal advice. Others pointed out that GDPR obliges companies to explain users' rights.
DHS bans DeepSeek, self-hosted included. A DHS order removes DeepSeek from federal civilian agencies (HN, Federal Register). HN readers noticed that the order covers "third-party products and services developed using DeepSeek implementations" and says the models "may contain malware or other security vulnerabilities," which bans local inference too. @jacquesm: open weights "are at least open to inspection ... the same can not be said about US providers." The obvious reply came quickly: "How do you inspect the weights for an intentionally hidden trigger that writes malware?"
Other Interesting Stuff
Anthropic's Public Benefit Fund. Anthropic is committing 5% of its shares to a new wholly owned LLC, subject to stockholder approval. The board can add 2% more "for every 50% increase in the company's value above a set baseline." The examples given are "large donations to traditional nonprofits or direct distributions to specific populations," with labor-market disruption named. None of the equity goes to lobbying or political giving. Add the co-founders' 80% pledge and the employee match, and "roughly 20% of Anthropic shares will be dedicated to the public good." The framing is pointed: "It is common practice for tech companies to offer escalating compensation awards to their executives. We are instead directing that equity to the public benefit." Separately, the WSJ ran "The obscure trust looming over Anthropic's IPO" (HN). HN noted how little the piece says about when the trust would act.
A Ruby maintainer quits. A longtime Ruby gem author archived every repository (HN, 460 points). The reasons: Rails' leadership, Ruby itself "now substantially developed using LLM agents," and Matz pitching Ruby as a "Programming Language for AI age." The author won't keep feeding "a commons that's being strip-mined by massive corporations to feed the very machines whose unthinking output pollutes it." The sign-off: "You want changes? Go ask your agent or something. I'm told that's what you're supposed to do now. I'm out." The HN comment worth reading came from @ttul, about a friend on a big project: the robots are good, "the joy is gone."
Data engineers wanted. swyx posted "Data engineers > AI engineers" and said demand for a "Data Intensive Applications" person over a "RAG guy" has never been starker in swyx's career. Latent Space is hiring at jobs@latent.space. Earlier, a month-by-month list of 2026's AI themes from Latent Space, built from AINews headlines for swyx's AIE Code keynote, is a quick refresher on the year.