The Loop Writes Its Own Goals, the npm Fight Armin Won't Finish & Rio's Borrowed SOTA

The /goal Loop Becomes the Unit of Work

Stop writing your own goal. The most-shared agentic post of the day, 268K views and 100+ replies, from Pietro Schirano: "I basically never write my own /goal anymore. I ask Codex to write one for itself, and one for each agent it spawns." The thread turned operational fast — Pablo Stanley distilled it into a paste-able template ("…write yourself a new goal and spawn agents in parallel — as many as needed… split the work into independent pieces, dispatch them concurrently, and synthesize the results as they return. Give each agent its own dedicated /goal."). Asked what to do when you don't even have the meta-goal, Schirano's answer was the whole philosophy: "All you need is a request." (And a practical footnote: he runs it on extra reasoning normally, dropping to low only to speed up the screen recording.)

LLMJunky's plain-English version. am.will reframed the mechanic for newcomers (for "$100+ plans only"): "/goal is essentially an agentic loop… start with small, well-defined goals. Have the agent build the goal prompt, steering it toward clear acceptance criteria and a way to test its work." His crisp definition of how /goal differs from /plan: "goal is an extension of plan — when the agent yields, it checks its work against the plan… if it's not done, it picks up where it needs to and works until it decides to yield again. Then it checks its own work to see whether the goal was met, and if not it keeps looping like this until the goal is met." The reply section reads like a status report on how people actually work now: AstrohackerLabs — "this is pretty much all I do now. Once you get it working for many hours at a time, or even days at a time, it starts to seem silly to do it any other way" — balanced by the honest caveat from Ryan: "small goals keep the agent honest. Vague goals produce impressive but broken code." OpenAI reportedly published a how-to on using /goal well, which one reader turned into a skill.

Satya supplies the enterprise theory. swyx surfaced the matching big-picture framing from Satya Nadella: "This is the first time we can create a real cognitive loop between people and digital systems… the real opportunity is not picking the best model but building a learning loop on top of models where human capital and token capital compound. You can offload a task, or even a job, but you can never offload your learning." His pitch — "build a frontier ecosystem, not just a frontier model… every organization owning the learning loop that encodes its institutional knowledge." The best reply deflated it perfectly: "the first cognitive loop between people and digital systems was me asking Stack Overflow the same question for 12 years."

What 24/7 loops actually look like. steipete put a face on the self-driving-codebase life (294K views, 117 replies): "Got a PayPal verification text and thought I'd been hacked, but it was just Codex signing up for a web service it needed." The replies are the joke and the warning at once — "things are getting serious if it buys a Netflix subscription to kill time while waiting for tasks to finish," to which steipete shot back "/watchparty." He also confirmed his crabbox harness is now portable: "it runs on Codex now. Or Copilot. Or our own harness."

Harness Friction: Pinning, Slop & Self-Signups

The npm fight Armin won't finish. The pinning saga from last week reached its bleak conclusion. Armin Ronacher, after a full day of it (19K views): "The deeper I dive into pinning deps for pi, the more I feel like actually using npm to install it is a fight not worth continuing. I cannot come up with a good strategy that allows deps to be pinned but still works as a library." The replies surfaced every workaround and Armin shot most of them down: bundle/vendor everything into one dependency-free package — but "the different packages all need to be installed independently, even just to support extensions; it's all very messy"; rewrite without deps (maybe even another language) — "that's not a realistic proposal at all." What he actually wants is small: a --locked flag on npm install that uses a bundled package.lock. The most fatalistic reply may be the truest — Andrei Chenchik: "pinning will not solve it. It's a problem of the volume and quality of dependencies."

"Root cause confirmed." Armin also named the agent-coding tic everyone now twitches at (190 likes): "Brother, if I get one more 'root cause confirmed' comment I am going to lose my shit." The replies piled on the genre — "Clankers love inventing fake root cause," "did you get more 'you are right'?" — the same week he admitted to burning two hours refactoring reload in pi so the agent could trigger it, and feeling like he was "making negative progress." The unglamorous texture of building the tools.

Codex-for-Linux gets a coat of paint. On the tooling side, am.will shipped a fresh build of the third-party Codex App for Linux (Arch, Debian, Fedora; github.com/am-will/codex-app) — rate-limit-reset banking, a developer mode for the internal browser, an activity profile with share cards, and an (untested) "migrate to Codex" path. The feature people zeroed in on: "rate-limit reset banking is the one Linux users are going to quietly love the most."

Panels, Pareto & the Fusion API Debate

Jerry Liu's sharp read on Fusion. OpenRouter's Fusion API (announced over the weekend) got its best analysis a day later. Jerry Liu: "This shows that frontier models alone do not own all the points on the cost-accuracy Pareto curve for knowledge-work tasks — in fact they may not be on the Pareto curve at all. The Pareto curve may be defined by a mixture of models, which any independent third party (e.g. an AI startup) has access to but the model labs do not." His kicker: the more specific your workflow, the more task-specific hill-climbing you can do — "that alpha is exploitable by any company that's not a frontier lab." The replies split cleanly. In favor: gamestoneai — "labs built the models and gave away the orchestration advantage. Classic"; Ferbin — "the economics flip the moment you route by task type." Skeptical: Philip C — "massive grain of salt. Do you really think OpenAI, Anthropic, even Meta/xAI would miss something so obvious? The benchmarks don't pass the smell test"; and the most substantive caution, from Saeed Anwar — "compound systems have compounding failure modes, so you're trading model predictability for ensemble accuracy, and that trade-off bites hard in regulated use cases." A couple of practitioners noted they'd already been doing multi-model code reviews and didn't even need a router — just a second LLM to critique the first.

An autoresearch benchmark with surprises. Zhengyao Jiang ran 7 frontier models across three autoresearch categories — ML engineering, harness/prompt engineering, and algorithmic discovery (45K views): "Fable-5 won overall even under cost constraint, but on ML engineering, the open model Kimi-K2.7-Code surpassed frontier models." The eyebrow-raiser was Opus 4.8 landing surprisingly low, which Jiang attributed to it being not quite as strong at "heuristic engineering and more conventional algorithm design." And the question that connects the two threads — would a "bag of models" win a verifiable domain like this? — drew a pointer to Sakana's AB-MCTS work. The mixture-of-models idea is clearly in the water.

Open Weights After the Ban

"Download intelligence while you can." With Fable 5 still benched, the sovereignty mood turned into practical advice. LLMJunky, 97K views: "Do yourself a favor. Stop what you're doing… even if you don't have a GPU, go download one of the latest local models and just keep it in storage. There may come a time when you can no longer access intelligence freely. 12–27B is enough." (The quoted model was Gemma 4 12B Coder.) The replies argued the premise both ways — the optimist: "anything out is out for good. They can't stop piracy, they can't stop open weights — torrents, binary newsgroups, IRC, etc." — and the realist: "it's not enough, but it might be all we get." LLMJunky's framing stayed defensive: "better safe than sorry. You never know what kind of dystopian world we're going to live in."

Rio's borrowed SOTA. The cautionary flip-side of the open-weight gold rush: the City of Rio's much-hyped "Rio 3.5" model, which had been making the rounds as a cheap municipal SOTA story, got forensically called out by Nex (699K views, 478 RTs): "Rio 3.5 ≈ 0.6 × Nex N2 Pro + 0.4 × Qwen 3.5… it even literally introduces itself as 'Nex N2 Pro' if you ask it without a system prompt." Their tone was "flattered, but attribution matters." The thread split between "that's how open source works — others can build on your models?" and people enjoying the "model sitcom" of it all — with one genuinely useful data point: IPlanRio reportedly spent R$500,000 ($100K) building the first generation, ~30× cheaper than an off-the-shelf system, per Exame. A merge is still a deployment; the embarrassment is the undisclosed recipe, not the reuse.

Also Worth a Look

  • Simon Willison's gallows relief. simonw: "I'm just glad nobody at the US government thought to try that Fable 5 'jailbreak' against Opus 4.x or GPT 5.x, or I wouldn't be getting anything useful done this weekend at all." The most concise summary of how narrowly the ban missed everyone else's workflow.
  • curl "euromaxxes." Armin Ronacher flagged Daniel Stenberg's post — "curl knows how to euromaxx" (daniel.haxx.se) — a small but on-theme note about European-rooted infrastructure the same week sovereignty was the headline.
  • Image models as a job-site tool. Outside the discourse, LLMJunky used ChatGPT to plan how to finish a section of poured sidewalk with limited reach — "not 100% perfect, but exactly what I needed, especially for leveling." A reminder that the most-used AI feature this week wasn't an agent loop, it was a multimodal model looking at a 2×4.
  • Knicks in 5. The NBA escape valve from last week resolved: the Finals are over, and Jerry Liu got his celebration ("Fable died. Knicks in 5!!!!"). LLMJunky used the moment for a sportsmanship subtweet about Wembanyama and Castle walking off without handshakes. Normal programming resumes once Fable does.