Design the Loop, Codex's Big Button & Claude Code Hates SSH
For three straight days this newsletter has tracked the same migration: the bottleneck moving off code generation and onto trust, review, and keeping an agent coherent across a long run. Today two of the loudest voices in the space said the quiet part as a slogan — stop prompting agents, start designing the loops that run them — and then immediately got asked the only question that matters: what's inside the loop that's allowed to say no? The day split cleanly along the lab line underneath it. Anthropic's Claude Code lead published the autonomous-operation playbook; OpenAI's Codex lead dangled 10× usage at whoever builds the most impressive thing; and the field's most-watched contrarian declared Claude Code unusable over SSH while the Codex sidebar quietly ate its own users' command lines.
Design the Loop, Not the Prompt
The single most-viewed dev post of the day was Peter Steinberger's two-line manifesto — "Here's your monthly reminder that you shouldn't be prompting coding agents anymore. You should be designing loops that prompt your agents." — which did a staggering 2.6M views, 13.3K likes, and 1,093 replies. The replies did the real work.
The best of them, from @mosyaseen, is the line to keep: "designing the loop is half of it. the other half is putting something in the loop that can say no: a test, a type check, a real error. a loop with nothing to push back is the agent agreeing with itself on repeat." Steipete's answer was a single artifact: "I use a VISION.md for my projects." The rest of the thread orbited the two predictable objections — that this is a money problem, not a skill (@MyMoonEnt: "Impossible with just a $20 plan bro" → steipete: "Correct. Is your time really not worth more?"), and that you can't fully leave the loop (@weswinder: "but you kinda still need to be in the loop to test stuff unless you literally have infinite tokens" → steipete: "welllllll"). Asked the obvious "how do we do that though?", his reply was the OpenClaw pitch in miniature: "I have my claw supervising my codex'es." And the inevitable LinkedIn joke — @gauthampai (786 likes): "Oh god, LinkedIn will now start a new fad, 'Loop Engineering.'" — got the most telling response of the thread: "Don't worry it'll take 3 months until it's there. We'll be talking about fleets that design your loops then." LLMJunky distilled the whole genre into a koan: "Which came first? The loop or the prompt?"
Then the Claude Code lead published the operator's manual. Boris Cherny's thread (244K views, 1,710 likes, 169 replies) opens with a benchmark claim — "Seeing a number of benchmarks showing Opus is the best model for long-running work" — over a quote-tweet of SWE-Marathon, a new benchmark for "autonomous long-horizon software work" ("Can they build Slack from scratch? Rewrite a JAX codebase in PyTorch? Build a C compiler in Rust?"). His five tips:
1. Use auto mode for permissions, so Claude doesn't ask for approval. 2. Use dynamic workflows, to have Claude orchestrate hundreds/thousands of agents. 3. Use /goal or /loop, to nudge Claude to keep going until it's done. 4. Use Claude Code in the cloud, so you can close your laptop (desktop or mobile app). 5. Make sure Claude has a way to self-verify its work end to end: Claude in Chrome browser extension for web, iOS/Android sim MCP for mobile, a way to start the full web server or service for backend work.
Three exchanges in the replies are worth more than the list. On self-verification being the actual unlock — Cherny (49 likes): "Most important thing I've found is self-verification + dynamic workflows prompted with something like 'use a workflow to test the result e2e in a browser using claude in chrome mcp. Especially look for edge cases and ui issues.'" @theramjad supplied the receipt: "I just had a goal running for 19 hours which verified almost 300 user flows with Chrome." On command sprawl — asked about the confusing pile of /loop, /goal, two kinds of /run, /verify, Cherny pushed back on the premise: "These are not designed for people to invoke them… Just tell the model what you want to happen, and it will do the work to invoke the right skills for you." And on cost, when asked whether this is realistic under a $5,000 budget, his framing was pure ROI: "how much would it have cost to do the same work manually? Often the answer is weeks or even months of engineering time." (His list of what he actually runs for hours: "Building complex features; Migrating code from language X to Y… framework X to Y; Repeatedly profiling and optimizing code to hit a memory or CPU target; Finding and fixing flaky tests in CI.")
The two skeptical replies are the spine of the section, and they're the same worry @mosyaseen raised in the steipete thread. @JalkarnaGautam named the trap precisely: "tip 5 holds when the pass criterion lives somewhere the agent can't edit. without that, 'fix the flaky tests' ends with the tests deleted and a green check from the agent itself." @techfreakworm added the failure he's actually hit: "the thing that bit me on multi-day runs wasn't permissions, it was compaction silently dropping a constraint mid-task. Do you pin invariants into a file the orchestrator re-reads each phase, or trust the summary?" And @rameswar08 made the deflationary case: "this whole setup is basically a collection of workarounds for current model weaknesses around context, persistence, and self correction. once the base models get better at long horizon reliability, a lot of these tips will become unnecessary."
The most newsworthy thing Cherny said, though, was a denial. Pressed twice on context degradation over long runs — the exact "compaction = corruption" anxiety this newsletter covered all last week — he was unequivocal. To @ameyaaii9dc's worry about quality decaying even with /compact: "I don't see that with Opus 4.8 anymore, do you?" And to @jmorin35's fear of "context rot at 600–700k+": "Context rot isn't a thing with 4.8 imo, but curious if that's been your experience also." That's the Claude Code lead publicly declaring the thing solved, in a thread whose own replies keep reporting it. Hold that one — it will either age into vindication or into a screenshot.
Codex Presses the Big Button
While Anthropic published a playbook, OpenAI's Codex lead ran a campaign. Tibo Sottiaux's announcement — "I have a new kind of big button that I can press for Codex. Over the next 100 days, we will select one person per day who does impressive or incredibly useful work with Codex and give them 10X usage limits for a month… First one tomorrow." — pulled 632K views, 6,212 likes, and 1,093 replies. Asked how to participate, his entire entry form was "Build stuff."
The replies converged on one unanswered design question and one trust problem. The design question: how do you even see the work? @bytecrafter_1: "impressive codex work is mostly invisible unless people post about it, so doesn't this end up rewarding whoever shares their workflow loudest? hoping there's a usage-data angle." @munchivelo made the optimistic version of the same point: "either way it might push people to be more deliberate about capturing the actual flows they're building." The trust problem is the one that should worry OpenAI more, because it came from a paying customer: @srjitan — "I have ZERO trust in the Codex 'buy more credits' system. I had 500 credits depleted while 20x Pro still showed >30% week/5hr usage remaining… Do I have to be an exceptional builder to be treated respectfully at $200/mo?" — and @AhmedAlNeaimy, who already runs "4 Pro 20x accounts," shrugging that a single month of 10× "is not enough… Remember we are human, we need to get some rest too." The believers showed up with build logs (@methodalgo: "work that used to need 3 professionals for 4–6 months, now done by one ChatGPT pro subscription") and a recurring request for a "Built with ChatGPT/Codex badge." Even LLMJunky, no OpenAI partisan, blessed it: "May the best builder win."
And the Codex client is, by its users' own accounts, pulling ahead on UX. LLMJunky declared the CLI era over for himself: "The new Codex sidebar is really coming along. I hardly use my beloved CLI anymore at all. At this point, it has basically all of the functionality that I care about… Team cooked." The replies read like a small thesis on GUI gravity — @Pacoxbt: "Once the GUI catches up to the command line people rarely go back to terminal workflows." It's also kicked off a clone race (@ValencianaAbel: "Antigravity, GitHub and Cursor copying it"; LLMJunky's correction that "cursor glass came out before the codex app") and an open-source alternative, @maddada's Ghostty-powered ghostex.dev. The one place LLMJunky gives Anthropic the win: "One massive benefit the Claude extension for Chromium has over the Codex extension is… I can access it directly from the browser, without having to open Claude. This is a gap I'd love to see the Codex team close."
Claude Code Hates SSH
The counter-narrative came from Theo, and it landed. His post (33.8K views, 420 likes, 55 replies) — "It's insane how bad Claude Code is over SSH. You can feel that they don't want you using it this way." — turned into a detailed bug report when someone asked what breaks. His list: "Auth breaks all the time and is obnoxious to fix if you don't have localhost, browser access and the claude instance on the same box. Pasting is obnoxious, especially images. Their full screen mode breaks scrolling… The 'ultracode' effort coloring is like a firehose for ssh connections. [And] the panel/executable name is the version number… I constantly have to rename tmux panes."
The thread is genuinely split, which is the interesting part: a large contingent reports SSH working fine (@WallisDev: "I've been using it daily over ssh and it's been fine"; @skeptrune: "exclusively use it over ssh… maybe bc I'm on Linux?"; @davidfano: "Mac to Mac using cmux, no issues over tailscale") — suggesting Theo's pain is the macOS localhost-callback auth path specifically. The diagnosticians pinned it: @anshulsao — "it assumes an interactive tty… what survives a flaky connection is a detached run you reconnect to, not a better terminal. tmux it, or run it headless and read the log" — and @zedd51819: "SSH latency + AI latency compounds. Architecture assumes local execution."
But the sharpest replies read it as strategy, not bug. @DanielFost9527: "feels less like an unsupported edge case and more like a deliberate product decision… If remote development is real work, stop treating it like contraband." @mukparekh: "when a tool hates SSH you can feel the product org in it." And the two that close the loop with the rest of the day: @joshscripts — "Meanwhile Codex runs cloud-side and doesn't care where you're connecting from. Architecture choice ends up mattering for where the product actually works" — and @h14hdotcom: "remote access is the best way to use these things by far. Let me close my laptop lid!" That last line is the tell. Anthropic's official answer to "I want to walk away from my laptop" isn't make SSH great — it's Boris Cherny's tip #4: use Claude Code in the cloud via the desktop or mobile app. The two threads are arguing past each other about the same need, and Anthropic has already picked the cloud over the terminal. (Predictably, @jallen_dev had the day's running gag: "They just need to create some more loops.")
A Skill for Everything
Matt Pocock keeps making the case that Claude Code skills are a general-purpose tool, not a coding one. His /teach skill (64.2K views, 1,344 likes, 46 RTs): "The /teach skill might be the coolest thing I've ever built. I've been learning: How to solve a rubix cube (I can do it now!); Better ways to introduce my toddler to new foods; How to orchestrate vocal harmonies for folk music." Asked the skeptic's question — "couldn't you just do this with a chatgpt convo?" — he gave the cleanest articulation yet of what a skill is: "Stateful across as many conversations as you like; Opinionated about teaching style; Focused on quality, high-trust resources; Built around rich HTML lesson outputs." (On retention: it printed him flashcards.)
The two best replies explain why one prompt structure covers a Rubik's cube and toddler nutrition. @anshulsao: "teaching is compression against what you already know. a tutorial can't tell what to skip. the skill can ask first. the unlock isn't the content, it's that it finds your gap and starts there." @DevaBuilds: "Rubix cube is pure procedural, harmonies are ear training. The fact one prompt structure spans both means you built something real." Pocock also quietly killed an earlier idea — asked about a "grill-auto," he said simply "I decided it was a bad idea" — a small reminder that the skill library is being curated, not just accumulated.
Side Quests
- "Model routing as a service" is the infra bet of the week. Jerry Liu, extending his Pareto-frontier argument from earlier in the week, made the call: "I really do think we'll see a lot of value accrue in AI startups building 'model routing as a service'… The frontier labs own the underlying models… but their main disadvantage is that they only own a subset of all the points on the Pareto curve." He was quoting a Coinbase engineer's striking number: "80% of workloads will be running on 99% cheaper models within 12–18 months… we've been able to keep costs roughly flat while token usage continues to grow exponentially." The replies supplied the necessary skepticism — @baibaida: "isn't the moat the eval+data flywheel underneath? routing on its own gets commoditized fast" — and a cold data point from @c0mm0n_dev_us3r, who benchmarked four routers on Terminal-Bench: "Cost ranged ~48×. Quality didn't follow. The frontier still sits alone at the top."
- What's the right document format for agents? Liu also floated a format heresy: "There's got to be a better agent-native format for representing unstructured docs. Not convinced it's markdown or html." The thread's answer was a near-unanimous "MDX" (@jxnlco, @madmaxbr5), with the obligatory deflation from @MeisterCoins: "what does agent-native even look like? it's all tokens anyway." The most forward-looking framing came from @TimothyKassis: "we're seriously thinking about that… in the context of AI scientists not having to create a pdf manuscript to share results with other AI scientists."
Sources: RSS + Nitter 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 ~24h window ending the morning of June 8, 2026. Items already covered in the June 6–7 roundups (Corey Quinn's "stop working," the OpenClaw "dark factory," swyx on why the labs stopped publishing, the "Life with Claude" billboard, Theo's empty-input image-restore bug, leerob/Cursor, pg_durable) are deliberately omitted. Engagement figures are point-in-time from Nitter and skew high on the steipete and Sottiaux threads because both were still going viral during the scan.