What the Swarm was reading — week of May 22, 2026
Seventeen posts this week, and an unusually high share of them snowballed into real arguments in the thread. The through-line was hard to miss: the agentic development era is no longer a forecast — it’s the substrate everything else is being argued on top of. What we read about tools, costs, culture, and even community politics now all sit inside that frame.
Here’s what came across.
The agentic SDLC moment, arriving from all sides
The headline of the week — Andrej Karpathy joining Anthropic — was the kind of news that lands as a punctuation mark, not a surprise. Puja noted that Karpathy himself has said he’d like to spend time at a frontier lab as an advisor and may not stay long, but the optics still count: from OpenAI, to Tesla, to Anthropic, the centre of gravity for the people thinking hardest about LLMs has shifted again. Jonas‘s read was the funniest one — “he saw /goal and took it literally” — and possibly also the most accurate.
The substance that made the week feel coherent came from Puja’s piece on the software inflection point — agents, OSS, and innersource. The argument is that agentic SDLC isn’t only changing how code gets written; it’s changing what “open” and “shared” mean inside a company. If an agent can read every internal repo at once, the wall between “our code” and “open source we depend on” gets thinner, and the question shifts from “what should we open source?” to “what capability gates do we encode, so the agent stays on the right side of the line?” Worth reading slowly. I read it multiple times. It’s worth it.
Sitting alongside that: a repackaged clip from Karpathy framed around “the death of vibe coding.” Timo rightly flagged it as a clickbait edit of a longer Sequoia interview (the original is here), but the underlying point is real and worth sitting with. Martin‘s reply was the most generous treatment of it I’ve seen: he’s been building a full KDE AI studio outside work as a way to understand what agentic development really looks like, and his summary is that we’re already past tinkering. The job now is “architectural guidance and UAT” — telling the agent what didn’t pass, why the approach was wrong, what you missed in planning. And that, he points out, requires a remarkable amount of documentation, role definitions for different agents (researcher, planner, dispatcher, implementer, reviewer, fixer), and patience. “Agentic development is a real skill that takes a lot of time, effort, and patience to get right.” That’s the line I’m keeping.
And then in a separate thread — kicked off by Oreoluwa sharing applied-llms.org — Timo delivered the cleanest one-paragraph reframe of where we actually are: “RAG is irrelevant these days. It’s all about context engineering and the harnesses do a lot of the heavy lifting.” Your prompts no longer have to be precise — the harness asks you the questions. Jonas piled on with the practical reading list: muster workflows, /context management, the Karpathy loop, the new /goal feature. If you wanted a single short reading list to catch someone up on where serious AI-assisted engineering sits in mid-2026, that thread is it.
The token bill comes due
Three posts that look unrelated until you stack them.
One: the OpenClaw founder burned $1.3M in tokens in 30 days, and admits that without fast mode it would have been 70% less. Timo‘s reaction was the operator’s one — he has OpenClaw running in a VM on his cluster, and “it just wastes a lot of tokens.” Fernando added the line worth quoting: tokens have a cost — not only economic, but environmental. The bragging tweet shows a real disconnect.
Two: Amazon is setting per-employee targets for AI token use. Marian caught the right Goodhart quote (”when a measure becomes a target, it ceases to be a good measure”), and that’s exactly the failure mode. The moment you incentivise tokens, you stop optimising for outcomes and start optimising for the wrapper around them. We’ll see a lot of this in the next year and almost none of it will produce better software.
Three: the Google / Antigravity pricing thread, where Martin walked through the absurdity. Gemini had 24h usage limits; Antigravity has 5h limits, but those aren’t the same as the app’s 5h limits (until they are). There’s a new €100/month plan if you need to upgrade, and additional tokens you can buy that may or may not roll over. Puja added the detail that lands the point: AI Pro now bundles YouTube Premium Lite, but only for the account holder, not the family. We live in a shitty future, indeed.
Stack the three together and the picture is clear: the economics of agent use are still being invented in public, and the incumbent platforms are reaching for every lever — token caps, pricing tiers, internal targets, bundled side-perks — to find one that holds. None of them feel stable yet. The companies that figure out where in the stack to extract value without breaking their own product loop will be the ones that survive this phase.
Supply chain, again, and an honest response
The week’s most useful security read was Grafana Labs’ update on the TanStack npm supply-chain ransomware incident. Quentin‘s pointer was the right one: it’s notable not for the incident itself — npm supply-chain attacks are now boring in their frequency — but for how upfront Grafana was about what happened. The bar for “transparent incident comms” keeps moving, and the companies that meet it without flinching are doing more for the ecosystem than the ones that try to spin.
Adjacent: Cloudflare’s cyber frontier models post. Marian flagged the strategy tips for multi-agent orchestration buried inside it — worth a read if you’re building anything where multiple agents need to cooperate without trampling each other. And Henning’s first LinkedIn post on Shadow Agents hit the same nerve from the enterprise side: the problem of AI agents quietly being spun up inside organisations by individual teams, with no governance, is the next iteration of shadow IT, and it’s already happening.
What AI does to a workplace
Marian brought a Fast Company piece on what increased AI-driven speed costs in terms of organisational alignment and culture. His own observation was the most concrete part: every time you ask an agent instead of a colleague, that’s one less personal interaction — and over enough such substitutions, the connective tissue of a company gets thinner.
Puja‘s pushback was sharp and worth quoting because it’s the steel-man for the other side: most of what the author calls “non-alignedness” is really just asynchronous work and growth. AI transcriptions don’t make people skip meetings — they just replace half-assed note-taking with something slightly better. And speed of execution doesn’t cause misalignment; it forces you to fix the gatekeeping processes that were hiding it. Both reads have merit. The honest answer is probably that AI raises the stakes on culture work, in both directions — speed makes good cultures faster, and bad ones fail faster too.
What I took from this week
Two threads, one shape.
One: agentic development isn’t the future, it’s the present. The Karpathy news, Puja’s blog, the applied-llms thread, Martin’s KDE-studio side project — these are all the same story told four ways. The skill that matters now is not writing the code, and not writing the prompt either. It’s structuring the context: documents, roles, gates, harnesses. The companies that figure out how to encode their actual taste and constraints into that structure will get exponentially more out of every model upgrade. The ones that don’t will keep buying tokens and producing slop.
Two: the token economy is in a transitional, slightly ugly phase. OpenClaw burns $1.3M and brags about it; Amazon makes spending tokens a KPI; Google reinvents its pricing tier weekly. None of these are equilibrium states. The interesting question isn’t which provider wins — it’s which abstraction layer captures the value once the dust settles. Smart money says it’s the one closest to the human, not the one closest to the GPU.
If any of these sent you somewhere interesting, reply and tell me where you ended up. The follow-on conversations are the part I’m here for.