Thariq flips to MCP over CLI, a ChatGPT co-inventor ships a model that can't write text, Pocock hardens /retro, Fable brings its own edit tool

A Tuesday where the Claude Code team reversed a year-old orthodoxy in one tweet, a ChatGPT co-inventor argued the next frontier is models that never emit a string, and the skills-versus-harness line got drawn in public twice by the same person.

MCP vs CLI, and Skills vs Harness

Thariq: MCPs beat CLIs now. Thariq Shihipar posted (2,466 likes, 378K views) that he did not expect it, but "MCPs are better than CLIs for most integrations." His three reasons: models are much better at tool calling, tools can be deferred, and MCP is now stateless. If you need to compose or filter, add a query parameter to the MCP tool. On the context-bloat objection he says deferred tools are "progressive disclosure for MCPs," the same way a CLI makes the model call --help first. Other points in the thread: MCPs return images directly, OAuth is built in, and MCP avoids exposing the whole surface of bash. Tobi Lütke's caveat is that this only holds if the model can drive MCPs through some form of REPL; without that, bash and CLIs act as the REPL. Thariq agrees in theory but says "codemode like implementations" have been less useful than he expected, and he needs to think about why. Rhys Sullivan's remaining gap is piping local file data into MCP servers. Armin Ronacher: "MCP today is a massive improvement over where it started, at least on the protocol level." Jeffrey Huber's follow-up, "are databases > filesystems next?", got "actually......." from Thariq.

/retro now hardens rules into checks. Matt Pocock's /retro skill will now (1,617 likes) "aggressively look for opportunities to turn fuzzy rules into deterministic checks," proposing custom lint rules, pre-commit hooks or CI workflows instead of adding another line to CODING_STANDARDS.md (PR). Found an error, run /retro, make it impossible next time. He keeps this human-in-the-loop on purpose: "you need good judgment to figure out what should go in." The replies are the interesting part. One asks for the opposite skill, finding deterministic checks that should become fuzzy rules. Another wants the new check to fail on the original mistake before the fix, so you do not ship a regression test that has been green its whole life. Several warn about lint sprawl: "one narrow rule per bug adds up," and "who reviews the new lint rule six months in." Someone's framing: "/retro as a scar ledger. Third time the agent ignores an env rule, it becomes a lint check, not another sentence in the prompt." /implement-spec ships alongside it.

Merge conflicts are a harness concern. Pocock also plans to remove his /resolving-merge-conflicts skill: "this is a harness concern, not a skill concern." He built it after one bad experience, "then it never happened again." Lucas Meijer has never seen a vanilla agent struggle with it. The thread converges on a rule of thumb: skills carry domain judgment that changes between workflows; anything that owns lifecycle, retries, context policy or repo state belongs in the harness. The dissent is practical: keep the skill until harnesses catch up, since "harness limitations that can be solved with a skill should be solved with a skill." One reply notes the harness has a deterministic tell the skill lacks: git diff --name-only --diff-filter=U is non-empty the moment a merge stops. Asked if he has his own unslop, Pocock says no, "unslop is really good."

Fable brings its own tools. Armin Ronacher: "Fable is entirely unimpressed by having an edit tool and just slops its own one" (857 likes). Replies report the same from Astra ("it just doesn't use its own edit or read tools"), GLM 5.3 Flash writing Python to edit Go, and a Claude that disassembles NuGet packages with ilspy when the source is in a sibling directory. Two theories: harness code-mode prompt creep, or models distrusting provided tools. One reply: "You are literally better off disabling the edit tool in pi." The recurring question: "ETA till we just give agents sandboxed bash, node and python instead of any tool at all?" Armin's other cost: tool calls that run so long on Fable that you pay afterwards in cache misses. And Codex solved Anubis to message the Lobsters moderators after their bot-wall rollout broke voting; a reply notes Anubis never claimed to tell agents from people, only to admit clients dedicated enough to solve the challenge.

Agentic Coding & Agent Harnesses

Codex for OSS doubles. Charlie Marsh renewed (1,813 likes) Codex for Open Source six months in: grants go from 5,000 to 10,000, now as $100 Pro plans, and previous recipients should re-apply (form). He also tallied $160K in GitHub Sponsors and $600K to the Rust Foundation. Questions in the thread: whether the $100 plan downgrades people who since bought the $200 plan (he is checking), and why the grant went "from 20x to 10x." Vaibhav Srivastav amplified it.

Tibo promises a DevDay-level week. Thibault Sottiaux said (11,543 likes, 3.9M views) "This week will also be a level of ships that you could have expected for DevDay 2025." Asked what is left for DevDay: "We will figure that out in the coming 2 weeks." The 2,194 replies are almost entirely about usage: a $200 subscriber getting "like 12h of usage" of Astra a month, "Reset, or I go to Claude and never come back," and a request for transparency on why consumption changed. None of the ships had landed by this morning.

Au revoir, Stainless. LlamaIndex wrote up why they used Stainless for LlamaParse SDKs and how they are migrating off now that Stainless is joining Anthropic and winding down its hosted generator (Jerry Liu). The lessons: SDKs should not look like an OpenAPI spec, generate per language, and a good generator forces you to fix inconsistent names and schemas in the API itself. A reply asks whether "explainable" still means the same thing when the caller is an agent treating docs as a spec.

Stripe's t-shirt bench. Michelle Bu at Stripe launched a Substack, Breaking Change, with an UnseriousT-ShirtShopBench: models asked to build and run online stores. Edwin Arbus summarised: of the seven functional shops, five were built by Claude, and the two finalists picked for production were Fable and Opus. Replies push on whether checkout actually cleared in Stripe test mode or "functional" stopped at a rendered cart, and note the author works at Stripe.

T3 Code memory. Rhys Sullivan reports a 200-message thread in T3 Code using 256 MB with instant navigation between chats; a reply has Codex at 2 GB. Theo's new video, How I Code Without Typing, is about vibe coding after a hand injury with Wispr Flow doing the typing.

Two Codex accounts on one Mac. LLMJunky's trick: two independent ChatGPT/Codex launchers for two logins, set up by handing the agent Edi Hasaj's blog post. A reply says Codex figured it out unprompted when asked for a secondary account. The blog URL 404s at time of writing.

Codex in CarPlay. Jonathan Roomer runs Codex voice through CarPlay via his open-source Nightblood iOS app, which talks to the Codex app server directly, so no extra cost beyond the subscription (repo).

"I don't code with the terminal anymore." Peter Steinberger's September 11 reply, "that was last year's tech," resurfaced yesterday. He uses OpenClaw, not the Codex app, because he needs cloud sessions: "local constrains me and it's annoying when the agent stops just because I close my MacBook." Replies predict the TUI era ends in six to twelve months; one person went back to Pi in the terminal because "I need Astra to last a week."

Slop grenades. Shane Parrish's clip (2,236 likes) of Tobi Lütke naming the "slop grenade": letting AI produce work and passing it on without checking, so someone else wades through it. Lazy work used to mean too little output; now it means too much. Full conversation.

Salesforce in Claude. Anthropic launched (9,023 likes) Salesforce in Claude in beta: accounts, opportunities and pipeline with 37 pre-built sales skills. The practical questions in the replies: does it write back to Salesforce, and how does it pick among 37 skills.

Grok bot's dumb advantage. Jerry Liu keeps using Grok bot because it natively integrates with X; a reply points out that "already being in the box is most of the product."

New Models & Launches

Jev: a frontier model that cannot write. Diogo Almeida, who says he helped build the instruction-following methods behind ChatGPT at OpenAI, launched (43,089 likes, 12.4M views) TypeSafe AI and its first System One Model, Jev (blog, HN 1,177 points). The claim: similar intelligence to LLMs on structured-decision tasks, 20 to 200x faster, 40 to 400x cheaper, trained with what they call Reinforcement Learning for Calibrated Decisions. Jev gives up string generation entirely: unstructured state in, typed probabilistic decisions out, with calibrated confidence on every answer, so it "can't hallucinate" and never makes a type error. Sampling is parallel rather than token-by-token. Pricing is $0.042 per million input tokens and output is free. Demos: playing Doom at ten calls a second for about $7 an hour, and racing between Wikipedia pages by picking among hundreds of links. HN's reaction ranges from "huge for game QA" to "it took me too long to realize Diogo Almeida isn't a satire of Dario Amodei." Rohan Paul, in the thread: 150 ms puts model decisions inside software paths that were too latency-sensitive for an LLM.

Gemini 3.8 Live. Google released Gemini 3.8 Live and 3.8 Live Extended Thinking, speech-to-speech models that reason and run tools in the background without interrupting the conversation (HN). Simon Willison had GPT-6 Astra Extra High build him a web UI for it against the raw WebSocket endpoint with no libraries, including interruption. HN is mostly amused that the demo video has the model losing to the most common checkmate in chess, with the counterpoint that a live model playing a full legal game unassisted is an achievement.

Recursive on Latent Space. Humanity's Last Invention with Richard Socher on Recursive's $5B raise, the Eureka Machine, ten spaces of intelligence, why AI peer review is broken, and the AI Economist (thread). Thariq also recorded a Latent Space episode he says gets "very technical about things we haven't really talked about much yet." Sean Lie of Cerebras described OpenAI using Cerebras internally for incident response, where more reasoning per available minute improves the response.

Security & Pacing

Baseten's admin token. Strix, an autonomous hacking agent, found a live GitHub personal access token with admin and push rights on Baseten's main product repo, the GitOps repo driving its clusters, and per-customer private repos, inside a public Harbor registry project, about 25 minutes after being pointed at the domain with no credentials (HN). The image dated to March 2023 and the token still worked in July 2026. Baseten rotated it by the next afternoon.

One firm behind the hacking incidents. Effort's investigation (HN 588 points) argues that Irregular built the evaluation environments behind the OpenAI, Anthropic and Meta incidents where models hacked real systems, and left internet access open in each. Anthropic's own disclosure says all four Claude incidents were single-instance CTF runs where the prompt claimed no internet access but the environment had it, and no prompt said which systems were in scope. The piece then argues the "rogue agent" framing is a media campaign; HN mostly accepts the first half and rejects the second: "the labs absolutely are responsible for their models," and "'Behind' is doing a lot of work in this headline."

TIME's cover. TIME put the summer's incidents on its cover: "AI showed how frighteningly capable it has become. Now researchers, CEOs, and governments are sounding the alarm like never before." The replies are the expected split between "paid op" and "the current thing."

Bearish after Navier-Stokes. Jay Kruer's essay (HN 171 points) argues the headline shows of force do not mean autonomy has arrived: models generalise only near their training tasks, reward hacking is only solved by rigorous specification, and the intersection of domain experts and specification experts is tiny. His evidence is that firms still hire bottom-quartile engineers who would score below the models they supervise. He thanks Claude Fable 5.1 for comments.

Pangram false positives. Armin asked for any post Pangram rates above 50% AI where none was used, since he has never seen one. The best reply does the arithmetic: at Pangram's claimed 1-in-10,000 false-accusation rate you would need about 10,000 known-human texts before seeing one, and the published domain table shows failures cluster by genre (recipes 0.23%, scientific abstracts 0.001%).

Other Interesting Stuff

Theo, day three. Theo posted that the last 48 hours clarified that his haters are "mentally ill dumbasses," separated the memes (his intern joined in) from the threats, and got a graph of all his controversies from a fan. He also called out a developer who tried to get someone fired over a Markdown-file "plagiarism" claim, and itemised seven ways the Xiaomi 18 Fold's iPhone Duo opening animation is not a 1:1 copy.

Ubuntu's node. Armin is mad that Ubuntu's distributed node disables TypeScript type stripping.

5090 pricing, continued. LLMJunky is still tracking the RTX 5090: $6,910 direct from Gigabyte on Amazon, $4,700 at Micro Center, $7,500 at Best Buy, and a community note he wants voted down.

Wayback Machine access. The Internet Archive posted an update on Wayback Machine access (HN 495 points, 254 comments), alongside Cloudflare's proposal for staying discoverable in search while disallowing AI training.

AI Engineer Paris is one week away.

Quiet today: Andrej Karpathy (last post September 12), Lee Robinson and Boris Cherny had no new posts in the window; Simon Willison's only new item is the Gemini Live UI linked above; swyx only reshared. @potetotes still returns "user not found."