The AGENTS.md Standards War, Cog's $10M Guarantee & Cloudflare Swallows Vite

After a week of pricing fights and framework eulogies, the conversation circled back to something more basic: what counts. What counts as a standard (a markdown filename, apparently). What counts as a passing eval (now measured in hours and dollars, with a $10M backstop). What counts as a usable open model when the job is a 100-hour agent run, not a benchmark sprint. Theo reignited the AGENTS.md-vs-CLAUDE.md war and accidentally made the better point about why the files diverge; Cognition put real money behind an eval number; and Cloudflare quietly bought the toolchain half the JS ecosystem builds on.

The AGENTS.md Standards War

Theo, who spent the last fortnight improbably defending Anthropic, turned on them over a filename. The post (76K views, 1,682 likes, 56 replies), quote-tweeting someone griping that "it's pretty annoying that Codex uses agents.md and Claude Code uses Claude.md… there should be some industry standards to this stuff": "There is a standard. It's Agents.md. Anthropic refuses to use the standard."

The genuinely interesting part was the tweet right before it, where Theo conceded why he keeps two files in the first place. His admission: "I've slowly come to appreciate that they are separate. Claude models need SO much handholding to behave correctly. I steer them entirely differently from OpenAI models, so I write my Claude.md entirely differently. It's also 4x longer and much dumber lol." That's the unresolved tension of the whole thread: a single shared file assumes the models want the same instructions, and Theo's own workflow says they emphatically don't.

The replies split into the three camps you'd predict:

  • "This is xkcd 927." @0x00_void: "xkcd 927 in real time." The most-liked version of the skeptic take, from @V3_Willy: "'standard' is doing a lot of work here. AGENTS.md is a convention popularized by one vendor. CLAUDE.md is the same convention with a different name from another. nobody convened, nobody voted, nobody owns it. the actual fix is symlinking one to the other in your repo until the dust settles."
  • "The filename is a red herring." The sharpest correction, from @18_priyansh: "the filename isn't the standard, the schema is. both parsers split on tool-call shape and skill links. naming the file is naming the fork twice. which spec wins on schema, agents or claude?" And @owenyuwono supplied the fact-check Theo skipped: "claude code reads agents.md as the fallback when there's no claude.md, so it's not refusing the standard, it just keeps a claude-specific file for the stuff the generic one flattens." (Theo's framing — "refuses" — is doing a lot of work.)
  • "Just symlink it." The pragmatist consensus, from @Farmasek: "the best thing we managed to do to unify, since our team uses different harnesses, is symlink claude.md to agent.md source of truth." @dahulilang put the realpolitik on it: "The standard that wins is the one that gets adopted by the most powerful ecosystem, not the technically superior one. AGENTS.md has better naming but Claude Code has better models. Until the market picks a winner, ln -s is the only real standard we have."

A few darker readings surfaced too — @denbvk: "I bet they think if somebody uses CLAUDE.md, it's a lock-in strategy to prevent users from churning to Codex" — and @metalagman_dev widened the complaint: "They don't support standard skills paths as well." The most deflationary take, from @SereneCorvidae: "It's honestly not gonna matter after a while. Markdown is markdown. You miss out on some harness optimizations, but modern models are already smart enough to interoperate between the two." Which, if true, makes the whole fight a transitional artifact — the file format equivalent of arguing about tabs vs spaces while the formatter quietly fixes both.

Evals Grow Teeth — Dollars, Hours & Guarantees

The day's most substantive launch was an eval, and it came with a financial guarantee attached. swyx announced (38K views) "the first eval ship from cog" and drew the contrast that matters: "@METR_Evals cap out at ~16 hours. Cog has private enterprise evals up to 100hrs, and is confident enough to put a financial guarantee on it." The datasets diverge in kind, not just length — METR's ground truth was "a combination of GPT-4o and GPT-5 to estimate human-equivalent times from compressed Claude Code transcripts" (34 sessions, 7 staff, rlog 0.83), while Cog's came from "258 sessions from 126 users across a diverse set of enterprise customers" doing real "java/typescript/python/c# feature dev, bugfixes, migrations" (rlog 0.74 on a held-out set). swyx called it "pioneering real world evals work and part 1 of a broader frontier code evals drop," with kudos to @annarmitchell and @ryanbai1412 for "the unglamorous last mile data collection."

The hook was the quote-tweeted Cognition pitch: *"AI should earn its keep. Introducing the AI Productivity Guarantee. If Devin delivers less engineering value than you're paying for, Cognition will fund your usage until it does, up to $10 million. It's time for the AI industry to stop maximizing tokens and start maximizing productive output."* After a week in which the running story was that near-free inference was a subsidy, a vendor putting its own money behind an output number is the natural next move — and the replies treated it as exactly that:

  • The cleanest framing, from @ankitships: "The 100 hour ceiling is the headline, but the financial guarantee is the real shift. A benchmark tells you it passed under known conditions. Underwriting tells you someone will pay when it fails under yours. Enterprises buy the second one."
  • The skeptic's math, from @NerdFuelDaily: "the cap is the tell. $10M total per customer (not per-failure) burns through in weeks at enterprise scale. underwriting math = P(claim) × min(damages, $10M) ≈ marketing line item. signaling, not insurance."
  • The market-pressure read, from @elalvarobalbin: "this guarantee just turned every lab into a 4-day reliability vendor overnight" — with @albe_sf noting a CloudBees report that "81% of leaders report production failures from AI-generated code," which is the demand the guarantee is priced against.
  • @subramanya named the catch: "this is the useful pressure. once pricing is tied to engineering value, review debt has to be visible too, not just hours saved."

The same "dollars, not benchmarks" thesis ran through swyx's other drops. The new Latent Space episode with Andon Labs is titled "Claude calls the FBI, AI CEOs, price cartels, Butter-Bench, & Luna" (latent.space/p/andon) — cofounders @lukaspet and @axelbacklund on "why dollar-denominated evals reveal what traditional benchmarks miss." Lukas's line, amplified by swyx: "Every day, it gets harder and harder to create tests that AI models can't beat. Reality is humanity's real last exam." And from Build, Sarah Guo's NoPriors × Latent Space sit-down with Satya Nadella (29K views) landed the framing that "$MSFT is at its heart still a tools company" with "big focus on agentic coding, harness & AI evals" — Nadella himself replied to thank them. The best reply, from @tsirbiladz3: "harness + evals is the real part imo. models get attention, but agentic coding becomes work software when the loop has tools, tests, logs, and reviewable evidence." @_virgil19 caught the tell: "'still a tools company' is the quiet tell. every lab is becoming one too, they just won't say it while the model is still supposed to be the moat."

Open Weights & the Long-Running Agent

NVIDIA shipped an open model designed for the exact thing the evals above measure — long agent runs. Mario Zechner (Pi's author) reported (25K views) — in a post amplified by Armin Ronacher — that "NVIDIA just released Nemotron 3 Ultra, a leading Hybrid Mamba-Transformer MoE open model for long-running agents. I've been test driving it these past few days, letting it work on Pi issues and it performed amazingly well!" Available via OpenRouter and his pi.dev. The architecture itself is the story — a Mamba-Transformer hybrid trades some attention for state-space layers that scale better over very long contexts, which is precisely what a multi-hour agent loop needs.

The replies were the usual mix of budget-king comparisons ("how does it compare to DS4F?") and one brutal anecdote from @bruce_x_offi that doubled as Pi marketing: he asked Claude Code (Sonnet 4.6 high) to add AMX-instruction CPU inference for a model and "1hr28min later — CC couldn't even figure out with huggingface token how to download the model properly, it got looped into partial download → re-download cycle… Same model in pi — already ds4 is running on cpu… I cannot use CC anymore, it's just horrible." One data point, but it's the kind that travels.

LLMJunky's parallel claim: the best Google model is the one you can download. His post (18K views, 328 likes): "Google's best models are, ironically, the open source Gemma models." Pushed on whether a Gemini 3.5 Flash price cut would change his mind, he held the line: "it wouldn't change my statement. The reality is that they are not just competing on the open source models, they're genuinely making some of the best small models available. That's the difference." The honest caveat came from @vibeparade: Gemma quality is great but "underwhelming on the agentic side… probably due to context window limits" — which is the same wall Nemotron's long-context architecture is trying to climb over. (Separately, LLMJunky noted Codex models got much better at Convex — "Code is law" — and shipped Remote Control for the Codex App on Linux.)

Cloudflare Swallows the Vite Toolchain

The biggest single tweet of the day by reach wasn't about agents at all. VoidZero announced (637K views, 4,508 likes, 676 RTs) it's "joining Cloudflare," with Evan You's team continuing to lead Vite, Vitest, Rolldown, Oxc and Vite+, all "remain[ing] MIT-licensed." The pitch: "bringing the Vite ecosystem and Cloudflare's platform even closer together." Armin Ronacher amplified it too — the JS-tooling consolidation is now a thing the systems crowd watches.

The reactions clustered around "tools as distribution":

  • @MaxHarlow_: "Cloudflare acquisitions work when tools become distribution: Vite users should see faster deploys, clearer observability, and fewer runtime handoffs."
  • @Bobchenjingbo: "the JS build layer getting faster and closer to where code runs helps everyone downstream."
  • The wary takes: @Tim_Damasc — "so now they have Void and Astro. what's the plan here?" — and the inevitable @Benchygo: "I think it's safe to say [the] js community is the most sellout out there."

The open question nobody answered: an MIT license protects the code, not the maintainers' attention. If the toolchain that powers Vite, Vitest and half the bundler ecosystem now optimizes for one runtime's edge, "even closer together" can cut two ways. It also rhymes uncomfortably with this week's earlier debate about who owns the dependencies you build on — except here the owner is a public infrastructure company, not a Linux distro.

Around the Ecosystem

  • Matt Pocock's AI Coding Dictionary is coming. He previewed (99K views, 1,394 likes) a glossary shipping later this month: "AI coding sounds complex (harness, model, agent, tool etc) but it's really not. You just need to understand the terms of engagement." The reply that nailed why it matters, from @curatedbyap: "The vocab is the barrier, not the difficulty. Most people 'bad at this' just never got told what a harness is." @ddavidddev: "this is the kind of thing that should be a pinned tweet in every AI coding tool repo." He's also floating a /grill-prep skill — "do a light grilling of your idea, then break out the areas of discussion into separate threads" — to fix the "I hit the dumb zone during grilling, help!" failure mode.
  • The home-cooked app, finally. Thariq resurfaced Robin Sloan's 2020 essay (42K views, 501 likes): "personal software was a bit early in 2020 but in 2026, it really can be as personal as a home cooked meal, or a handwritten letter" (robinsloan.com). The reply that captured the shift, from @mosesxu: "i'm building little personal tools constantly now, ones that were never worth a saas seat or a whole weekend. the shift isn't only that building got cheap. it's that bending myself to someone else's app finally got more expensive than making my own."
  • Lee Robinson keeps pushing back on the 'builder' merge. His thread (135K views, 1,326 likes) rejects the podcast line that engineering, product and design are "all merging into a 'builder' role": "someone still has to wake up and care 100x more about Product or Design than anyone else. It is their Main Thing." The reply that one-upped it, from @nichochar: "You're in for a bunch of pain if engineers ship a bunch of slop, let alone non engineers."
  • Codex love from an unlikely corner. Steipete amplified Christoph Nakazawa's post — "I've been dumping on OpenAI with low effort meme tweets… but Codex is the best DevX acceleration product of all time" (cpojer.net/posts/modern-engineering-values) — a notable concession in a week otherwise dominated by Claude-vs-Codex steering complaints (see the top section).
  • The mysterious 347. A fun LLMJunky observation: ask a model to "pick a number between 1-500" and, "much like humans tend towards the number 37, the model will almost always include the mysterious number, with 347 being the most common variant." Training-data echoes leaking into "randomness" — a small reminder that these things don't pick numbers, they recall them.

Sources: RSS + thread scans of the accounts in TASK.md (Pocock, Theo, Thariq, LLMJunky, Mitsuhiko, bcherny, Steipete, swyx, Simon Willison, Karpathy, Jerry Liu, potetotes, Lee Robinson) for the ~24–36h window ending June 5, 2026. Engagement figures are point-in-time from Nitter.