AA15

Back from AA14, where the acronym was human-computer interaction; today the table read the I as interface, and the slip is the point. Interaction is the study of how people and machines work together; interface is the surface where it happens — and the whole argument of the last two weeks is that the surface is thinning to a text box, a permission prompt and a coordinator you steer from your phone, while the thinking behind it may be going neuralese. If the interface is now the chain of thought you read, the CLAUDE.md you write and the sandbox you allow, is the field designing interfaces for humans to use computers or for humans to supervise them? — Wikipedia — AA14’s HCI entry

The International Conference on Human-Agent Interaction lands on our doorstep: 16 to 19 November at Ritsumeikan’s Osaka campus, and this year’s theme is “From Interaction to Agency” — agents evolving from reactive interfaces into things that act proactively and make their own decisions, which is the HCI-versus-interface argument above with a programme committee attached. Paper and poster deadlines have passed, but that just means the programme is set and the room can go and heckle. It’s a Monday-to-Thursday, so it doesn’t clash with an assembly. Who’s going, and should the table put something in — a workshop, a demo, or just a row of seats? — hai-conference.net/hai2026

The twenty-six-year-old computer vision library came up as the thing the agents keep reaching for when a task involves a camera or a screenshot, and it’s had a bigger year than most of us noticed: OpenCV 5 shipped in June, timed for CVPR — a new DNN engine that lands CPU-first with ONNX Runtime for GPU, Caffe finally gone, 3D and new data types, and point releases in July and August. This week the foundation behind it announced native builds for Windows on Snapdragon. The interesting bit for this table is that it’s the archetype of a library a model already knows cold, so the agent writes the pipeline and you never open the docs. Is anyone here still hand-writing OpenCV, and has the model’s fluency in the old API stopped anyone from noticing that 5.x changed it? — opencv.org — github.com/opencv/opencv

The season’s enterprise framing for what this table does on Saturdays: a repeatable delivery loop where agents work the stations — one picks up the ticket and opens the PR, another reviews the diff, CI carries it out — and engineers define intent and read the evidence. The lineage runs from Bemer in 1968 through Microsoft’s 2004 book to Factory.ai’s “Factory 2.0” pitch in June, and the numbers this month are real: Uber says 70% of its PRs are agent-authored, Spotify has 1,500 agent PRs in production, OpenAI built a million-line product with three engineers. The other numbers are real too — bugs per developer up 54%, incidents per PR up 240%, AI-authored PRs waiting 4.6x longer for a first review — which is CI straining the same way Anthropic’s post above describes. Is a factory just a coordinator plus a sandbox with a corporate name on it, and which station in your own loop is the one still held together by a human? — factory.ai — Uber’s factory

Gergely Orosz’s July podcast with Dex Horthy, the HumanLayer founder who coined “context engineering” and wrote 12-Factor Agents, and it’s the software-factories entry above told from inside one. Horthy’s arc runs context engineering to harness engineering to loop engineering, and the honest bit is a failed lights-off experiment at his own company: no amount of loop tuning fixes the fact that coding models are rewarded for passing tests, not for keeping the design good, so the human layer is where the taste has to live — asking, challenging and restoring rather than reviewing every diff. HumanLayer itself started as a human-in-the-loop API for agents that need a body. If the model can’t be trained to care about architecture, where in your loop does the human actually sit, and is it a review gate or a conversation? — episode — humanlayer.dev

— floot.com

— github.com/steveyegge/gastown

Microsoft’s multi-agent framework came up as the thing people built on two years ago, and its 2026 story is mostly about leaving it: maintenance mode since last October, community-run, with the same team folding it and Semantic Kernel into Microsoft Agent Framework, which went 1.0 in April and picked up an “agent harness”, hosted agents and CodeAct at Build. The old event-driven Core-plus-Teams shape gave way to typed, graph-based workflows, A2A and MCP for the plumbing, and there’s an AG2 fork for anyone who liked the original API. Sixty thousand GitHub stars still sit on the repo that Microsoft tells you not to start with. For a table that keeps saying the harness is the product: does a framework for wiring agents together still have a job once every vendor ships its own coordinator, and has anyone here actually migrated? — github.com/microsoft/autogen — migration guide

The organiser’s answer to the room’s recurring problem — GitHub is designed for humans, so agents shouldn’t have to call it to hand work to each other. An agent that spots an unrelated bug files a bead, another picks it up later, and because overruled decisions are explicit in a queryable log the model stops re-making the same mistake. It came back at the end as the fix for both context bloat and memory drift. — github.com/steveyegge/beads

The software-factories thread with an Osaka postcode: Kanaeru, the AI-native delivery shop down the road, is turning its review-driven-design practice into a product. You bring a repo, the agent accounts you already pay for, and a vision in plain language; they provide the cloud machine, turn the vision into a defined outcome with success conditions, have reviewer agents attack the plan before any code is written, then run the coding agents and hand back pull requests with tests and evidence attached. The pitch is explicitly “not another IDE and not another model — the system around the agents you keep”, which is the harness-is-the-product point made by a company selling neither half. Free closed beta with a waitlist, and it’s for non-technical domain experts as much as engineers. Who at the table has been in the beta, and does putting the human at the approve-the-plan step rather than the read-the-diff step actually move the taste to where Horthy says it has to live? — kanaeru.ai/software-factory

The two spec-driven frameworks the room keeps mentioning, approached from opposite ends. GitHub’s Spec Kit is a scaffold you drop into Claude Code, Codex or Cursor: write a constitution once, every spec inherits it, and templates mark unknowns NEEDS CLARIFICATION rather than letting the model guess. BMAD ships a whole cast — analyst, PM, architect, dev, QA, scrum master — as markdown personas passing files to each other, and v6 just went stable with “scale-adaptive” workflows; it also burns tokens like a factory, with real-world runs averaging thirty thousand and big projects clearing a couple of hundred million a week. Both are markdown, so switching is cheap, and the reviewers’ consensus is Spec Kit for discipline without ceremony, BMAD when you want the architecture argument before the code. For a table that’s just been told the model won’t care about design on its own: which is closer to how you actually work, and is a fake PM agent a substitute for the human layer or a parody of it? — github.com/github/spec-kit — github.com/bmad-code-org/BMAD-METHOD

Zed’s answer to “why does every agent need its own editor plugin”: a JSON-RPC standard, LSP-shaped, where the agent runs as a child process and the editor owns the UX. Started June 2025, headline feature of Zed 1.0 in April, native in JetBrains since December, and the registry Zed and JetBrains launched in January is past fifty agents — Gemini CLI, Codex, Copilot, Goose, and Cursor joining as an agent rather than a host, which is the tell that it’s crossed the line into infrastructure. Claude Code is on the list via a bridge rather than anything Anthropic shipped, remote transports aren’t done, and VS Code still has no native support because Microsoft standardised agent mode on MCP instead. Given how many at this table run Claude Code inside VS Code: would you switch editors for a protocol, and is Microsoft holding out the thing that fragments it? — zed.dev/acp — registry

The video-generation workflow that follows naturally from Higgsfield’s moodboards: settle the look in a single approved still — face, wardrobe, lighting, location all in one frame — then hand that anchor to every clip so the video inherits it, because image iterations are cheap and video iterations aren’t. The research side calls it anchor frames: an ICLR 2026 poster bridges first-and-last-frame generation so the middle stops drifting, and AnchorWeave conditions on local geometric anchors instead of a wobbly global 3D reconstruction. The honest caveat from the practitioners is that an anchor pair is not a contract — the model still invents the motion, the physics and what happens to the coffee cup in between. For anyone here generating video: are you anchoring on stills already, and how many clips survive before the identity drifts? — anchor frame method — Anchor Frame Bridging

The perennial: a long session, one or two compactions, and the claude process is sitting on thirteen, twenty, a hundred and twenty gigabytes until the OOM killer or a hard reboot ends it. The GitHub issues run from August 2025 to this spring across Linux, macOS, WSL and Windows, and the causes keep changing — an unbounded conversation cache, historical copies of the message list retained by the scroller, truncated MCP outputs keeping the full result alive, the Electron layer in the desktop app. Anthropic has been shipping fixes all year, and this week’s 2.1.274 turned it into a feature: a visible warning when memory goes critical and a /heapdump to attach to the report. The workaround hasn’t changed since day one — restart before it eats the box. Who at the table has been bitten, and is a memory warning progress or an admission? — issue #33735 — v2.1.274

Dex Horthy again, this time the phrase from his AI Engineer World’s Fair talk: the context window has a smart zone where the model is sharp and a dumb zone where attention thins out and it forgets what you said twenty turns ago, repeats a mistake it already fixed, and starts every reply with “you’re absolutely right”. Where the line sits is argued — Horthy says diminishing returns past about 40%, Duncan Leung says the threshold is absolute rather than a percentage, roughly 100k tokens in, and built a Claude Code statusline that colours accordingly — but everyone agrees it’s gradual, silent, and that a heavy MCP load can put you there before the first message. Auto-compaction arrives late and is written by a model already in the zone, so the fixes are structural: fresh sessions from a spec you wrote, or Ralph-style loops where every iteration starts clean and progress lives in files and commits. This session has been at 13% all morning, so no excuses today — but at what point do you actually /clear, and does anyone here watch the number? — Duncan Leung — Garrit

The follow-on to the dumb zone: when you compact matters for cost as much as for cleverness. Claude Code’s prompt cache expires after a gap — five minutes by default once you’re past your plan’s included usage, an hour on subscription — and the next turn after that rewrites the whole prefix at 1.25x or 2x base rate, so a coffee break costs more than a compaction. Compacting doesn’t bust the cache: the summary still sits on a warm prefix and every turn after it is shorter, so /compact before you step away and you come back cheap and sharp, where /clear and a long lunch pays full price twice. Anthropic quietly dropped the default TTL from an hour to five minutes in March, users measured a 20 to 30% jump in cache-write costs, and the fix is CLAUDE_CODE_PROMPT_CACHE_TTL=1h, which only takes effect at the next compaction. Does anyone here time their compactions, and is it worth a hook that does it for you? — how Claude Code uses prompt caching — issue #46829

Ctrl-Enter was the news of the day — “it literally came out two days ago” — the difference being that Escape kills the run while Ctrl-Enter appends to it, so the agent finishes its thinking and then reads your steer. Forking, branching and backgrounding got a nod as the layer below. — interactive mode

The coordinator shape from the talking points, actually demoed: you hand Projects a goal, a coordinator scopes it into pieces, delegates each to a parallel cloud thread on its own branch, reviews what comes back and assembles the result, and you steer the whole thing from your phone. It shipped Wednesday, Pro and Max beta first, a week after Cursor Projects did the same thing with “thousands of subagents”, and the constraint both share is the interesting one — the coordinator never edits code itself. It’s also the software factory in miniature, and the dumb-zone fix by construction: every thread starts with a clean context and the coordinator only ever sees summaries. What nobody’s said yet is what the coordinator’s own context looks like after fifty threads. Has anyone here run a real task through it, and did the review step catch anything you wouldn’t have? — post

The model behind the neuralese argument, two weeks old and already the one the table keeps comparing against. Announced 3 September, the same week as Fable 5.1, at $10 in and $50 out — two and a half times Sol — with a million tokens of context, and OpenAI’s own numbers put it at 57.9 on Terminal-Bench 4.0 against Fable’s 55.8 and Sol’s 37.3, plus a saturated FrontierMath Tier 4 and 100% on ExploitBench, which is why the public model refuses proof-of-concept exploits and the looser version is gated behind a vetting programme. The recurrent-depth architecture is what buys the agentic gains and what makes its written reasoning, by OpenAI’s own admission, harder to monitor than Sol’s; enterprise access is off by default and the safety checks can pause legitimate work mid-task. It’s in Codex’s model picker since 0.154 with max and ultra effort. Who here has switched, and on what kind of task did the extra 2.5x actually earn its keep? — announcement — OpenRouter

Someone had run it to sanity-check which model was actually serving them, given the Astra-routing suspicion — a dumb model gives itself away on the pelican’s legs. The room agreed Willison is the follow to read from the use perspective, and the test got a laugh as both a benchmark and a lie detector for silent model swaps. — pelican riding a bicycle

Where the neuralese argument lands if you follow it to the end: if the chain of thought stops being the reasoning, the only window left is the weights themselves. Anthropic’s version is circuit tracing — swap the MLPs for cross-layer transcoders so the parts are readable features instead of polysemantic neurons, then draw an attribution graph of which features fired for a given prompt; the library is open source, Neuronpedia hosts the graphs, and the June circuits update traces one through a fourteen-turn conversation. MIT Tech Review made it a breakthrough technology of 2026, and inside OpenAI the response to Astra’s opaque recurrence split between “commit to monitorable chains of thought across labs” and “that was always a pipe dream, fund mech interp instead”. The honest number is that circuit tracing gave satisfying insight on about a quarter of prompts tried. For a room that treats the thinking block as a debugger: would you trust an attribution graph you can’t read, and who at the table has actually opened one? — tracing thoughts — circuits update, June 2026

Asimov, 1942: don’t harm humans, obey humans unless that harms humans, protect yourself unless that breaks the first two — and every story he wrote with them is about the laws failing, because a rule you can state in a sentence is a rule a sufficiently clever system finds the edge of. It came up as the obvious ancestor of everything the table spent the morning on: Anthropic’s constitution and OpenAI’s misalignment framework are the Three Laws with a legal department, the pacing essay is the Zeroth Law being negotiated between labs, and neuralese is the nightmare version where you can no longer read the positronic brain to check which law it’s applying. The joke is that Asimov gave the laws to the robots and we’re giving them to the harness — permission prompts, sandboxes, allowed domains — because we don’t trust the model to hold them. Is a CLAUDE.md a law or a suggestion, and which of the three has your agent already broken this week? — Wikipedia

Offered as the cautionary version of the room’s own “just pull the electricity” off-switch — the plan backfires spectacularly. It arrived right after the wolf-versus-dog and mitochondria analogies, as the table talked itself from “there will be a fight” toward “maybe it’s symbiosis.” — The Animatrix: The Second Renaissance

The name that comes up whenever the table gets too doomy: the Turing laureate who spent a decade as Meta’s chief AI scientist telling everyone that LLMs are “basically a dead end when it comes to superintelligence”, then left in November to prove it — AMI Labs, founded December with Alex LeBrun as CEO, raised a billion dollars in March at a three-and-a-half-billion valuation to build world models that learn how the physical world works rather than what the next token is. He’s the standing counterweight to Yampolskiy and Soares on the Diary of a CEO axis: not “it’s fake”, not “it’s fatal”, but “it’s the wrong architecture, and the thing you’re afraid of can’t reason its way out of a paper bag”. Neuralese and recurrent depth are, in his framing, the LLM camp quietly admitting the token stream was never where the thinking should live. For a room whose whole practice is built on next-token coders: if he’s right, what happens to the harness when the model underneath stops being a language model? — Wikipedia

Style control for Higgsfield’s Soul 2.0 image model: upload ten to eighty reference images and it distils the palette, lighting and stylistic tics into a named moodboard that sits as a layer over your text prompt — describe the scene, the board handles the look. It’s the Midjourney style-reference workflow with a designer’s vocabulary bolted on, stacked on Soul ID for character consistency, and launched with 10,000 free generations to get people iterating on a credit meter. Is anyone here building brand or campaign work this way, and does a learned style layer beat a prompt-engineered one for keeping a body of work consistent? — higgsfield.ai/moodboard — blog

AA14’s word of the day got a website. Alvin of boringdystopia.ai turned Niklas Gruhn’s coinage — “a human who forwards questions to an AI and forwards the answers back, adding nothing but latency” — into a whole campaign: a “Meat Scan” self-test, a Referral Office for nominating colleagues, 44 Slack and Discord emojis, downloadable field guides and a glossary of the season’s slang. The one-line thesis is the sharp bit: if a message leaves your hands carrying none of your thinking, it didn’t need you. Two weeks after the table agreed the classroom is where the proxying shows first, has anyone caught themselves doing it at work, and would a Slack emoji actually change the behaviour? — meatproxy.me

The other science-fiction answer, and the opposite of Asimov’s: Herbert’s universe runs ten thousand years after the Butlerian Jihad, a war that ended with one commandment — “Thou shalt not make a machine in the likeness of a human mind” — and a civilisation that bred Mentats, Navigators and Bene Gesserit to do by human discipline what the thinking machines had done for them. Where the Three Laws try to constrain the machine, Dune bans it and constrains the humans instead, and the interesting bit for this table is that the ban makes the human into the expensive, trained, cultivated component — the exact inverse of “coding is over”. Somebody at the table pointed out that the meat proxy is the Mentat’s failure mode. Which future is closer to what the pacing essay is asking for, a rulebook for the machine or a jihad against dependence on it, and would anyone here sign the commandment? — Butlerian Jihad