Felix retires the coding interview, Claude goes into drug development, AA v5 widens the open-weights gap, Codex resets again, and physicists take custody of OpenAI's Yang-Mills proof

Agentic Coding & Agent Harnesses

Thariq's AI-native dock, and forks that keep merging upstream

Thariq's personal project got the biggest reaction of the day among the tracked accounts (8.5K likes, 596K views): "an AI-native macOS dock with custom modes (writing, building, etc.) to help me stay focused". The agent "is able to customize mode behavior and add new modes, it uses jev to classify my activity + notifications in real time & act on them." Thariq isn't sharing it yet: "there are still many edge cases (would not recommend using)", but "this is the kind of HCI work I've been wanting to do for a while, but messing with OS internals has always been very time consuming". ThePrimeagen replied "bro, just come to omarchy" and Thariq answered "I like my Mac". @jxnlco's "Oh cool I didn't know you were technical" got "barely".

Thariq then asked for "a new way to distribute opensource software such that people can modify via agents but also get upstream automatically merged in", because "coding agents can do this but they won't do it automatically" (2.6K likes). Replies pointed to existing attempts. Kieran Klaassen linked ghuntley.com/remix, Fabian mentioned Theo's PATCH.md, and dex cited the Kubernetes kustomize design and a "design-iterated-loop" skill for keeping a fork current "while preserving our changes". Steve Ruiz said tldraw had "considered shipping migrations as skills but ended up building around it". The skeptic's case came from @someRandomDev5: "Upstream changes may inherently conflict with what you're building (and this is often even EXACTLY the reason why you forked it in the first place)."

Matt Pocock wants a word for tasks agents can't get wrong

"I feel like I need a word for the type of task that agents are really good at" (1.2K likes, 456 replies): tasks "where it has 100% of the data, it can run experiments, verify independently, and hill-climb towards its destination … On these tasks my trust level is totally maxed-out". Mario Zechner's "agent-complete" was the top reply (187 likes). Philo Groves suggested "Verifiable", noting that "Many tasks may seem unverifiable right now, until we build the verifiers." Matt leaned that way with a caveat: "some tasks are partially verifiable, which is where agents get into trouble thinking they're verified them when it's actually half-finished. I'm thinking like '100% verifiable'." Other candidates: "Legible", "Known", "claudable", "latchkey" tasks, and "they shall be pocock tasks".

Earlier, Matt posted an example of domain knowledge paying off (1.4K likes). While building a video editor, Matt knew "certain consonants can fiddle around with silence detection" and told the agent. "The word 'fricative' is enough for it to start busting out full IPA … Turns out, knowing things is really fucking valuable." The feature is an LLM pass "over the deterministic detection to edit for rhythm and cadence." Andy Wojcicki took it as an argument against the "cavemen" skills: "Precise language is there for a reason".

OpenCode 2.0 hits 1M daily users, and dax on open-model demand

dax: "OpenCode 2.0 now has 1M daily active users … that happened in 2 weeks" (1.2K likes). Asked for more "fluff", dax replied with a mock press release whose real news is that "Phase 3 of Operation Cheepseek" gives "$80 of usage for your $10 a month subscription" for DeepSeek v4.1 Flash.

The bigger thread was about demand. Gavin Baker had cited DeepSeek "monetizing at $38 per GPU hour", and dax wrote "i just don't see where the demand for this is today … whenever we talk to providers it's clear all of their other customers are 100x smaller." Then the numbers (2.4K likes, 390K views): "openrouter is the most successful model router, you can almost serve 100% of their deepseek traffic yesterday on a single GB300 rack … this is worth 6-7M...even if there's 10x this demand secretly that's still < 100M of compute." Pushed back on, dax added that "agentic coding produces 100x the volume as any other usecase … even so, multiply my numbers by 10, it's still tiny", and that "a lot of it must go straight to deepseek the company". The conclusion: "someone's gotta make the apps". In the meantime dax is still buying GB300 capacity ("if you have capacity you're not using we will take it").

Two more dax posts drew replies. "it's getting close to 2 years of coding agents being a thing … the majority of discussion on here on where they're going, what's cutting edge, what matters has been primarily wrong … you had to actively ignore what was going on here to win in this space", with a follow-up: "every day since we started i woke up and read messages saying 'OpenCode isn't doing X they are going to die'". And on API design: "i used to put a lot of effort in making them beautiful because human eyes had to consume them … now i care more about explicitness even if its ugly".

OpenCode shipped too: plugins for OpenCode Desktop, where "one package can ship server, tui, and desktop plugins", and a database explorer. @jlongster described a new OpenCode project "for building high-quality cross-platform native apps easily. we're all committing directly to it through ai", including work to "perfectly match macOS font rendering".

Theo: no dashboards, and sell to my agents

"If I have to go to a dashboard to use your product, I'm not going to use your product. If anything, I'm more likely to tell Claude to rebuild your product." (1.6K likes). Joe Jordan's ""Contact us for pricing" lol no" was the top reply, and Valentin Ignatev admitted: "My first gut reaction was 'holy larp', but after … reflecting upon all my AWS experience I'm afraid to admit I agree with you." In a reply to Corey Quinn, Theo added: "I like my agents, I don't want yours. My agents need more tools, more data, more capabilities. Sell me things I can give my agents. Don't sell me a worse agent." Corey agreed: "I'm not going to stuff my entire business context into your crapbot".

Other T3 Code news:

Quick hits

Three essays on working with agents (HN)

  • AI Promised Less Work. I Have Never Worked More. (HN, 631 points, 628 comments). Orestis Ioannou, VP of Technology at Bold.org, charted GitHub contributions since 2019: back at full-stack-engineer pace, with 2,102 contributions by October 27, August 2026 the busiest month ever (427), and 16 days of 24+ contributions. "Agents did not just make my existing workload faster to finish. They made it easy to give myself a much bigger one." Top commenter sailfast: "It's the best time ever to build. That excitement lasted about a year - but now it's a treadmill I can't escape".
  • The Mythical Man-Month in the age of AI (HN, 452 points). Jan van den Berg argues Brooks's Law "mutated": "nine agents still can't deliver a baby in one month", and "instead of managing people, you are now managing agents." The most-discussed reply, from rimeice: "it gets you the wrong answer so fast it's now the best way to figure out the right requirements."
  • Here be monsters: What do you still need to know when agents do the work? (HN, 168 points). The argument is for fundamentals and breadth so you know when to dive in. The author, adrq, replying in the comments: "you still need a baseline to be able to tell when the map you've been given looks right but isn't".

Personal Agents

am.will reviews Claude Duet, then takes half of it back

am.will's first review of Claude Duet called it "the least useful personal agent" for now. The complaints: it's a separate app, it can read cloud conversations and Claude Code sessions (with Remote Control on) "but it cannot interact with those threads", and it seemed missing from the desktop app. The praise: it runs on Opus 5.5, it tracks tasks and suggests new ones, it shares plugins (including Link/Shop Pay), and it is "quite proactive". Six hours later: "You CAN access Duet from the app. It was hidden in the 3-dot menu! This changes my review totally." Alex Colón described the intended hierarchy as "you > Duet > channels > projects > cowork/code sessions", which "takes away some freedom (by design, a la Anthropic's remove subagents guidance) but simplifies management. Voice mode is terrible".

Grok Bot: hundreds of PRs a day

Lauren (@poteto) wrote that on the Grok Bot team "engineers, designers, and PMs all contribute multiple hundreds of PRs a day", made possible by "making our code friendly for agents, by making the easy thing the right thing" (1.8K likes). Lauren called it "the best example today in the industry of where all those tokens are going", since "software development is an iceberg - new features are only a small fraction of the work in maintaining high quality software." Lauren also shared tibot, the bot Lauren uses to watch X mentions and keywords, send one sorted digest a day and block trolls ("never posts for you"). Grok Bot 0.77.0 added permission settings for auto-review rules and multi-question cards.

Grok Bot also had outside fans. In a clip shared by Fireside Alpha, Gavin Baker said that "Grok Bot is the most impactful AI product to me and to the firm that we've had yet. I glued together all this stuff with Claude Code", and that Baker is "up to like 17 Grok Bots". SpaceXSI's sales team described Account Intel, Quote and Deal bots (x.ai writeup), and @mntruell said "we're beginning to hire Bots onto our team." On the other side, dax said of something @ryanvogel teased with "it's almost time": "our team has been using this for months now and it's probably why no personal assistant agent product has stuck for us".

Claude & Anthropic

Felix Rieseberg retires the coding interview

"The coding interview is dead, at least for my team at Anthropic. I've seen enough to know that it no longer provides signal. I removed it from loops this month." (8.3K likes, 876K views). The accompanying post, Hunger beats tenure, says "There's no more point in giving you a little React challenge than there is in quizzing people about assembly; Claude is now better at both", and that newer engineers now "fly past people who are probably better coders and have been around much longer." The advice for engineers "without social capital, status, or power": pick business impact over backlog, pick a radical idea, build it over a weekend, focus on one good idea, and lean into a fresh perspective. Felix's example: two relatively new engineers, Anthony Morris and Alice Zhao, built a working prototype of a graphical Claude Code over one weekend, and "The first public version shipped roughly eight days later." Felix hasn't found a replacement yet and is "scanning resumes and portfolios for little prototypes that exhibit taste, empathy, and principled thinking."

Lauren (@poteto) replied "love it! thanks for leading the way" and Thariq "++". The doubters: "Most companies and organizations aren't anthropic and will be doing them 7 years from now lmao". Jeremy Bernier said they did "a take-home coding test for Anthropic last week" and almost wished for the traditional interview, "b/c then at least I'd know what to expect".

Claude in drug development: six partners and its own programs

Anthropic published early results from research partnerships with BigHat Biosciences, Bristol Myers Squibb, Divergence Labs, Enveda, Generate Biomedicines and The Column Group (7.9K likes, 1.7M views). It has also started its own drug discovery work, "including programs for rare childhood forms of muscular dystrophy, where Claude handles much of the scientific and operational work." The case studies:

  • Generate Biomedicines: six scientists used Claude Code to redesign part of a CAR-T therapy. One design was "about 70% shorter than the standard version" and let immune cells make "roughly 10 times as many receptors".
  • Divergence Labs (food allergies): Claude "largely autonomously, generated roughly 35,000 antibody designs in about a month".
  • BigHat: Claude, connected to the lab's data system, proposed changes that the lab turned into about 1,010 variants in about 10 weeks. Partway through, Claude and the Anthropic researcher decided the standard heat-stability and clumping metrics were misleading and proposed different ones.
  • Enveda: over 1,061 pages of manufacturing documentation, Claude "caught all 11 issues that human reviewers had spotted—and flagged seven they'd missed", and cut review time by 89%. A regulatory document draft took four hours instead of "about three months by hand".
  • The Column Group is starting Pleiades, a company that will use Claude to run more drug programs than usual for a small team. BMS is using Claude to estimate how likely an early-stage trial is to succeed.

On HN (211 points), people questioned the line "we need to work with many more biotechnology firms and integrate Claude more deeply into their development processes" ("the wording here feels very self-serving"). enraged_camel's explanation for why biology moves slower than math: "the verification loops are much longer in applied sciences."

Also from Anthropic

  • 40% off for government. A new OneGov agreement with GSA gives federal, state, local, tribal and territorial governments 40% off Claude Enterprise, Claude for Government and the API, on "every product surface" including cloud providers in the coming months (tweet). Anthropic's examples include a state IT developer who "moved nearly 900 API methods to a new framework in two days" and a federal team that used Claude Mythos to find a takeover bug "A human penetration test the month before had missed". Top reply: "40% off, then it goes through a contractor who bills it back to them at 3x lmao".
  • Stanford professors' books. The Stanford Daily reported that Tobias Wolff and Chang-rae Lee opted out of the $1.5B Bartz settlement and filed Cruz v. Anthropic in May, seeking up to $150,000 per work instead of the roughly $3,000 in Bartz. On HN (513 points), the submitter changed the title after commenters noted that the ruling held that training on legally acquired books is fair use, and Simon Willison noted that Anthropic "claimed that those pirated books had not been used to train their models" and has since scanned "vast numbers of second hand books".
  • Feedback keeps the whole conversation. A privacy center page hit HN (218 points): thumbs-up/down feedback stores "the entire related conversation … for up to 5 years". Commenters asked whether the "How is Claude doing" prompt in Claude Code counts.
  • Latent Space on the Claude Platform. How Anthropic ships Claude with Claude: Separating the Brain from the Hands (video and transcript) has swyx interviewing Katelyn Lesse and Angela Jiang, who lead engineering and product for the Claude Platform. Chapters cover the Opus 5.5 launch and pricing, Managed Agents, memory and dreaming, CPU constraints, rebuilding the Messages API in Rust, and protecting credentials in sandboxes (tweet).
  • An S-1 reader. Jerry Liu's team parsed all 318 pages of the S-1 into a reader built with LlamaParse. Gavin Baker called the filing "pretty clearly an upside surprise".
  • Apple II pelican. Benj Edwards posted that Apple II Agent (Opus 5.5) drew "a pretty good pelican riding a bicycle" on the Apple II. Simon Willison: "These ones are pretty great!"

Codex & OpenAI

Tibo's Day 23: another reset

"Day 23/ Much requested iCloud Main & Calendar plugins. Just kidding, today is a reset. Still much requested though!" (6.5K likes, 864K views, over 1,000 replies). The iCloud Mail and Calendar plugins did ship in ChatGPT the same day, on all plans. The top reply asked "why is our subscription burning so much quicker now? It seems a lot more than 2-4x quicker." The next asked "can we just do this until the new year?" Tibo said it would land by midnight PST, then: "And it's done. See you tomorrow for some good news." (3.2K likes). The replies guessed what that might be, from "a model beating Opus 5.5" to the full 20x limits coming back.

Physicists take custody of OpenAI's Yang-Mills proof

OpenAI's claimed proof of the existence of quantum chromodynamics (covered yesterday) now has a custodian. KITP posted a statement from the committee that will handle it: Michael R. Douglas, David J. Gross (chair), Arthur Jaffe, Igor R. Klebanov and Nikita Nekrasov. On October 16, OpenAI asked Gross "to assume custody of the proof, decide on the best way to release it to the community, and shepherd the release process", and the committee says it has "no responsibility to OpenAI." The statement calls the proof of quantum Yang-Mills "including coupling to fermions" "an important step towards putting QFT on a rigorous mathematical footing" and credits Tadeusz Balaban's renormalization-group work from the 1980s. But "confinement and the mass gap remain very important open challenges", which, as @aiSignalDrift put it, means "not the full Millennium problem yet". The committee will first check that the Lean formalization's definitions and problem statement match standard physics and mathematics, release the Lean proof after that, and publish "the full 1147-page LLM proofs" with its assessment.

Several replies objected to the withholding. Francesco Maggi wrote that "Once a public claim of a major scientific discovery is made, the evidence supporting that claim should be made public", and @aiSignalDrift asked "Why not publish the Lean theorem statement now so anyone can audit what was actually claimed?"

Models & Benchmarks

Artificial Analysis v5: private tests, and the open-weights gap grows to 12 points

Artificial Analysis released Intelligence Index v5 (3.2K likes, 569K views), with "60% of the Index … now based on evaluations with private test sets or held-out answers." New additions are Cognition's private FrontierCode 2.0 ("graded on whether a repository maintainer would merge the agent's change") and Terminal-Bench-Science. HLE-Diamond replaces Humanity's Last Exam, and GDPval-AA, SciCode and AA-LCR are gone. Claude Opus 5.5 (max) leads at 58, ahead of GPT-6 Astra (57) and GPT-6.1 Sol (55), with Gemini 4 Argon next at 50. The open-weights gap went from 3 points to 12. Kimi K3.1 Pro (weights not yet released) scores 46, and the best released weights, DeepSeek V4.1 Flash 1017 and GLM-5.3, score 32, below Claude Haiku 5.5 at 36. GPT-6.1 Terra was the cheapest top-10 model at $0.96 per task, and Fable 5.1 the most expensive at $13.04.

Many replies didn't believe the ranking. The top one: "sonnet 5.5 and 6.1 terra being above fable 5.1 here doesn't exactly give me confidence". Others asked "majority private now. so we just take your word on the scores?" and "Where's the med/high/xhigh leaderboard??" The tracked accounts mostly agreed with the gap:

Also in models

  • GPT-6.1 Terra on ParseBench. Jerry Liu found Terra "spikes at table parsing", fifth overall on ParseBench, ahead of 6 Sol and behind Opus 5.5 and Astra, but "struggles on layout, charts, and formatting", at about 1 cent per page.
  • LFM3. Liquid AI released an open-weight family of six ~2B models (text, vision, audio, omni and two decision models) on one backbone pre-trained on 22T tokens, with 256K context and a text model that "fits in about 1 GB at 4-bit precision" (HN).

Other Interesting Stuff

  • The Classification of Finite Simple Groups, in Lean. Axiom formalized the first-generation proof "with every logical step checked by Lean's kernel" and found "counterexamples to a closure assertion used in a published route to the Unbalanced Group Theorem", which a new paper works around (HN, 203 points). pfdietz: "The formalization is 43.1M lines of Lean."
  • Genie 4. Google DeepMind announced its next world model, now generating audio and dialogue with the world (HN, 857 points, 502 comments). The top threads were about cloud-streamed games as DRM and whether Genie 4 responds to inputs worse than Genie 3. One commenter noticed that a bust dropped on a watermelon "should've toppled over, if the physics were properly modeled".
  • Trump wants to ban the term "artificial intelligence". Trump said on Tuesday that the administration would "officially" ban the term "in all relevant documents and everything else" in favor of "super intelligence", per CNET, building on a late-September executive order for federal agencies (HN).
  • Wikipedia's AI cleanup. IEEE Spectrum covered the WikiProject that removes AI-generated writing (HN, 475 points).
  • Tailscale SSH update. After the tailworm audit, Tailscale published interim guidance: set TS_SSH_DISABLE_FORWARDING=1 to mitigate "several of the more serious findings" until patched clients ship (HN, 265 points). Commenters noted that headscale users are affected too, since the bug is in tailscaled.
  • Survival LLM. A local assistant plus emergency library that runs on your own computer (HN, 318 points).