Five Archetypes for the Melted Org, the Model-Picker Paradox & Touch-Grass Coding

The Five Archetypes of the Melted Org

The day's runaway thread came from the person who'd know: the creator of Claude Code, reflecting on what jobs even are anymore. Boris Cherny (496 replies, 1,163 RTs, 10,769 likes, 1.08M views): "As engineering, product, design, DS, etc. melt into a new kind of role, I was reflecting on what roles might look like in the future." Looking at the Claude Code team, he saw five archetypes — and crucially, ones that cut across job titles:

  1. Prototyper: comes up with brand new ideas; churns out many, most of which don't ship
  2. Builder: quickly turns a prototype/idea into production-grade product/infra
  3. Sweeper: cleans up the UI, simplifies the code and system, unships, optimizes performance
  4. Grower: takes a built product and iterates to improve Product-Market Fit
  5. Maintainer: owns a mature system to make it secure, reliable, fast, efficient as it scales

The payload was the un-bundling of role from function: "these roles are not really tied to job function — eg. across Anthropic, some designers match category 1, some 2, some 3; same for engineers, PM, DS." His staffing model is by product stage, not department: "A product that is new and pre-PMF needs people strong at 1+2+3 … strong PMF needs 3+4+5 and some 2. Maybe product roles of the future will look more like this, and less like the domain-specific roles of today?"

The sharpest reply warned against the frame itself. Kun Chen (350 likes): "i never quite liked the idea of defining archetypes, because it's easy for people to look at it and think 'ah that's who i am' and never question themselves again. … when i'm starting a new project i'm a prototyper and builder, but very soon i become a sweeper … then a grower and maintainer. if i box myself into a particular role, i would have to give up the project somewhere along the way. … stay flexible and be obsessed with whatever is important for achieving your goal, and care less about the boundaries of roles that will continue to blur over time." Cherny conceded immediately: "Totally agree. Roles often change over time/project."

The blunt question underneath it all came from Pachu (29 likes): "Why do we need builder and sweeper roles if coding is mostly solved? Can't you just loop Claude on it?" Cherny's answer was a useful tell on where the harness actually is today — reply (19 likes): "Claude can help with all of these to varying extents, and will improve over time. Agree that Sweeper/Builder Claude is quite good at today." (Read between the lines: Prototyper, Grower, and Maintainer are where the humans still earn their keep.) Gerrard Lipscombe named the structural shift: "these are effectively broken down into stages of a business/product rather than isolated verticals/departments. It requires people to be more generalist, and I guess that's precisely what you're saying is now possible, hence the melting." It's the same throughline as this week's role-collapse discourse, but stated from the eye of the storm: the org chart isn't getting flatter, it's getting fluid.

Too Many Knobs - The Model-Picker Paradox

If roles are melting, the model menus are doing the opposite — fractally multiplying. Theo (94 replies, 32 RTs, 1,884 likes, 287K views), quote-tweeting a complaint about GPT-5.6's Sol/Terra/Luna tiers ("how do you choose between Sol low vs Terra xhigh"), one-upped it with Claude's combinatorics:

  • Fable 5 (low, mid, high, xhigh, max, ultracode)
  • Opus 4.8 (low, mid, high, xhigh, max, ultracode)
  • Opus 4.8 [1m context] (low, mid, high, xhigh, max, ultracode)
  • Sonnet 4.6 (low, mid, high, xhigh, max)
  • Sonnet 4.6 [1m context] (low, mid, high, xhigh, max)
  • Haiku 4.5 (low, mid, high, xhigh, max) "And don't forget about no reasoning modes as well as fast mode for Opus."

The most-liked reply was a scream into the void. noel (72 likes): "i hate this reasoning effort juggling. just fucking train the fucking model to decide how much it has to think for a question/task." Pavlo (45 likes): "I think this is the best answer." The same instinct showed up as resignation — Paulo Mateus: "I just use everything in high and whatever it does is what it is. If the AI does not know enough to choose, I won't bother!" — and as a consolidation wishlist from KYDNO: "Haiku needs to be killed, 1M context needs to become more cost efficient then become the default … deleate one of the reasoning options (xhigh or max prolly) and then we chilling."

Even the experts can't keep the matrix straight. Peter Steinberger (50 likes), adding a row Theo missed: "you forgot adaptive thinking that they introduced in 4.6 🫠." And Theo drew the one durable distinction in the mess — on verbosity: "I think of 'verbosity' less as a model setting and more as a user preference, similar to memory or personality settings." The thread is funny, but it's also the cleanest statement of a real product failure: the labs have offloaded their hardest scheduling decision — how much compute this specific task deserves — onto a dropdown, and users are unanimously asking for it back.

The Dumb Zone & the Order of Skills

Matt Pocock spent the day turning context-management folklore into something you can feel. Pocock (53 replies, 14 RTs, 664 likes, 54K views): "Here's a way to get a visceral sense of the dumb zone (starts at ~120K tokens on SOTA models): 1. Agent does something dumb. 2. /rewind to before the stupid thing. 3. /compact. 4. Run it again. Watch it get smarter." The mechanism, in his words (33 likes): "Agents get dumber as they go along because of attention degradation." On the fuzzy threshold — Pocock: "40% of the context window was for 200K windows, so around 80K. Everyone agrees one exists, but it radically changes per harness/model/task."

The live debate was whether auto-compaction has fixed this. Kent C. Dodds: "How do you feel about automatic compaction? I think cursor's must be great because I rarely feel like the model gets stupider as conversation length grows." Pocock's caution was specific and battle-tested — reply (13 likes): "/compact is still too unreliable because you need to be super careful not to do it mid-phase. A mid-phase /compact can make the agent radically lose track of its task — it's hit me almost every time I've tried it." The takeaway isn't anti-compaction; it's that when you compact is now a skill in itself.

The companion thread was a pop quiz on harness internals. Pocock (29 replies, 82 likes, 30K views): "Imagine I have /skill-1 and /skill-2, both with disable-model-invocation: true. I tell Claude Code: '/skill-1, /skill-2 do XYZ'. Which of the skills will be invoked?" The answer was a small horror story about churn — Pocock's reveal: "~3 weeks ago: /skill-1 only. ~1 week ago: /skill-1 and skill-2. Today: /skill-1 only. IMO invoking both is the correct behavior" — i.e. the behavior has flip-flopped across three Claude Code releases. The cross-harness color was the useful part: Tak: "codex (at the harness level) will actually expand the skill file into the user input when invoked this way," and Theoklitos Bampouris (9 likes) issued the portability warning: "Please don't make your skills Claude specific. disable-model-invocation: true works only with Claude." As skills become the unit of reuse, their invocation semantics are quietly becoming an interop problem.

Touch Grass - The Remote-Agent Command Center

Yesterday Theo predicted devs would move their agents off the laptop within six months. Today the receipt arrived, and it had a vendor name on it. Hamel Husain (61 replies, 19 RTs, 508 likes, 76K views), RT'd by steipete: "Codex desktop is miles ahead of Claude for remote access via mobile or another computer. Codex allows you to see all running sessions on all devices from any other device, including mobile. You can truly go touch grass with Codex. Brings me joy every day."

The Claude counterpoint exists but loses on ergonomics. Deep Thrill: "Claude Code's /remote-control from the cli lets me control my local session from the Claude mobile app. Does codex have similar functionality I've missed?" Hamel: "Yeah but codex just makes everything visible without requiring you to do any slash commands. It's just built for remote way better — you basically have a command center for everything that could be running." The "is this even healthy" subplot wrote itself: Ken Warner (8 likes): "I've been sitting in the tub the last 3 hours doing a full codebase audit, triage, planning and implementation handoff to subagents all using voice mode from my phone," to which Hamel could only reply: "idk if that's healthy 😅 maybe get dry." The infra fight from yesterday's roundup has a frontrunner this morning: not whichever lab has the smartest model, but whichever one made the fleet of running agents legible from a phone.

Restriction vs. Open Weights

The week's access-policy story stopped being one argument and split into three fronts.

Front one — the establishment turns. Mark Kretschmann (27 replies, 34 RTs, 254 likes, 14.9K views) surfaced former White House AI czar David Sacks against Washington's restriction push, citing a WSJ "nightmare scenario" of Chinese models catching up in cybersecurity while US policy debates self-restraint: "America does not win the AI race by slowing itself down. It wins by building faster, scaling infrastructure, lowering energy costs, and exporting … Restricting American AI may feel like safety. But if it accelerates Chinese alternatives, it is not safety. It is surrender with extra paperwork." The second-order caveat, from Jan Stevens: "slowing your own frontier models only works if everyone else slows down too."

Front two — the villain origin story. LLMJunky (12 replies, 71 likes) named names: "Dario is sure doing his best work trying to get open source models outlawed. He's going to keep pushing for this." His follow-up: "It's honestly getting my blood pressure high … The audacity." Dan B supplied the cynic's summary: "Create the problem, sell the solution."

Front three — Europe's non-option. Armin Ronacher (16 replies, 132 likes, 11K views): "I happen to think it would be a good idea for the EU to approach Anthropic. On the other hand it's a severe misreading to believe that Anthropic would in any way align with the EU. It's unfortunately quite laughable :(" His diagnosis — reply: "the problem is less the EU than the environment Anthropic is in. It's both my investments and strategic interest caught by the US government." The replies converged on the only real answer: _cjk: "EU/UK need their own Anthropic. The parallel is France and nuclear." Florian Brand (15 likes): "best for the EU would be to build out DCs like crazy and bankroll mistral (and/or fund competing national labs)."

And the builder counter-move: make the whole question moot by going local. Ahmad Osman (138 replies, 124 RTs, 1,792 likes, 264K views): "MASSIVE NEWS. Teamed up with NVIDIA to make Local AI The Default" — a banner being unfurled at AI Engineer World's Fair this week, with Mike Bradley supplying the rallying cry: "Opensource AI Must Win." When the frontier pulls behind glass and the regulators reach for the open-weights tier too, the consistent reflex all week has been the same: route around it, and increasingly, route home.

Also Worth a Look

Did coding agents just rewrite the math on legacy codebases? Thariq (39 replies, 598 likes, 180K views), on Riot announcing League of Legends Classic (old champions/runes/items on the original map): "this has to be because coding agents change the engineering math on how it is to work with or port a legacy codebase, right? anyone at Riot able to confirm?" Ray Amjad (26 likes) added the asset angle: "combines well with image models getting better — put in the old game assets and generate higher-res scaled-up versions in seconds." And Rhys Sullivan (38 likes) had the line of the thread: "the reason we're not entering a fast take off is because people keep porting games from their childhood to play and are getting distracted." Thariq: "as planned." (WoW Classic, AoE2, and Diablo II: Resurrected got name-checked as the pre-AI precedents — but the speed is the new variable.)

AI Engineer World's Fair opens, fully sold out. swyx confirmed the event crossed into sold-out territory — leadership track, workshops, and all late-bird tickets gone, with 65 free side events across SF — then reported from the floor: "we registered 1,000 people today … tmr and tuesday going to be absolutely batshit." His most candid note was on the hardest track to curate — swyx: "because i'm not a design engineer myself, this track is one of the harder ones … very fortunate to befriend Geoff who lent a hand." Expect the next few roundups to be heavy with dispatches.

"You can just make things. Literally." A small, wholesome counter-current to the policy gloom — LLMJunky (14 likes): "here's the story of how I let Codex decorate my son's birthday party." The lone reply was a useful reality check from Screenclutter: "this … sort of minimizes how much effort you've put in to verify each stage of the pipeline. You don't one-shot a party decoration because your kid deserves it." The "you can just do things" energy is real; so is the verification tax under it.

Agents as accidental accessibility allies. A reframe worth filing, from Tak (56 likes, 9.9K views), RT'd by steipete: "AI agents may end up doing more for web accessibility than accessibility laws. Agents rely on accessibility APIs to use computers and websites efficiently. As websites adapt for agents, people with disabilities benefit too." Weston Johnson named it: "Reverse curb cutting." The dissent was sharp, too — tymrtn: "for sites where publishers (journals, news) are trying to block unlicensed agents entirely, we may end up hamstringing accessibility in the process."