Payment Required

What we talked about

tell a fable

Anthropic’s Claude Fable 5 sits at the top of Cursor’s CursorBench — 72.9%, eight points clear of the field — but the leaderboard isn’t the whole story. Fable 5 was pulled once already and then restored, and BridgeMind now claims the brought-back July-1st build came back quietly nerfed: on their own BridgeBench, debugging cratered 86.2 → 25.9, refactoring 73.6 → 38.4, and hallucination got worse, 75.9 → 61.7. Their theory — new guardrails are tripping on ordinary tasks and silently falling back to Opus 4.8, so “this is not the model that got banned,” and Anthropic “owes everyone an explanation.” Open question — when the model behind an API can be swapped, guard-railed or throttled under you with no version bump, what is a headline benchmark score actually worth, and how would you even catch a silent nerf on your own workload? — CursorBench evals — BridgeMind — Fable 5 came back nerfed

Phones as an ad hoc mic array

The table riffed on a build — pool everyone’s Android phone into one distributed microphone, each catching a speaker from a different spot, then transcribe and stitch the streams into a single “who said what” record. Turns out it has a name — an ad hoc microphone array — and both Microsoft Research and Google have chased it: the spatial spread lets you pick the cleanest signal per speaker and do real diarisation instead of the concatenated mush one phone gives you. The bit the room kept circling back to is clock-sync across devices; channel selection, ASR and merging are mostly solved, the alignment is the hard part. Open question — is a table full of phones a genuinely better meeting recorder than one decent mic, or does the sync headache eat the whole win? — Microsoft Research — ad hoc mic arrays from personal devices — Google Research — group conversations via sound localization

Higgsfield

The AI-video one — Higgsfield pitches itself as a full creative workspace rather than a single model, aggregating the frontier video generators (Sora 2, Kling 3.0, Veo 3.1, Seedance) behind one interface so you can prompt them all and compare side by side. Its calling card is camera control — the Cinema Studio mode simulates real optics (pick a lens, focal length, depth of field) and locks characters across shots, closer to a film set than a text box. The bit for this room is the MCP server: point Claude or another agent at it and generation becomes just another tool call, orchestrating the best model per shot without you learning each one — a neat callback to last week’s WordPress-MCP thread. Open question — does a multi-model aggregator with agent hooks beat going direct to Sora or Veo, or is it a nicer wrapper around the same drift-and-break limits? — higgsfield.ai — Higgsfield MCP for Claude

Seedance 2.0

Straight off the Higgsfield chat — Seedance 2.0 is ByteDance’s frontier video generator, one of the models Higgsfield aggregates, and the interesting bit is how far past the text box it goes: feed it up to 9 images, 3 videos and 3 audio files alongside the prompt, and it lifts motion, camera moves and character style from your references rather than making you describe them in words. Character consistency across frames, video extension, built-in audio, credit-based pricing. It’s the reference-driven counterpoint to Cinema Studio’s virtual-lens approach — show it, don’t tell it. Open question — is prompting with reference footage the interface video gen actually wanted all along, and does “bring your own motion” finally make output art-directable rather than a slot machine? — seedance2.ai

Remotion

The deterministic end of the video thread — where Higgsfield and Seedance above generate pixels, Remotion programs them: video as React components driven by props and a frame counter, rendered to MP4 with a headless browser. A returning topic from AA04, and it slots differently this week — as the layer an agent can actually diff, test and re-render, versus generation you can only re-roll. Data-driven explainers, per-user clips, charts that animate from a JSON feed — all just code an LLM can write. Open question — as the gen models get art-directable with reference footage, does the code-first approach keep its edge for anything beyond data-driven templates, or do the two ends meet in the middle? — remotion.dev — github.com/remotion-dev/remotion

Headroom

The token-bill thread — Headroom is an open-source compression layer that sits between your agent and the model, stripping the redundant bulk agents drag through the context window: duplicate file re-reads, full API responses when you needed one field, verbose log dumps. Under the hood it routes content through purpose-built compressors (AST-aware for code, SmartCrusher for JSON) rather than lossy summarisation, and the claimed numbers are eye-watering — 92% fewer tokens on SRE debugging, 73% on issue triage — with 43k+ GitHub stars behind it. The write-up bills it as “the Netflix tool,” though the repo itself sits under an individual account. Open question — with subagents now burning tokens in the background by default, is a compression proxy the missing layer of the agent stack, or does silently rewriting the model’s context risk deleting the one detail it needed? — Headroom — Medium — github.com/chopratejas/headroom

Coze

ByteDance’s second appearance of the morning — after Seedance the room looked at Coze, their no-code/low-code agent-building platform: drag-and-drop workflows with LLM, code and condition nodes, plugin and knowledge-base management, RAG built in. The interesting move is that the core got open-sourced as Coze Studio (plus Coze Loop for debugging, evals and monitoring) under Apache 2.0 — Go backend, React front, runs on 2 cores and 4GB, so a self-hosted agent factory is genuinely on the table. It’s the visual counterpoint to this room’s CLI-first habits. Open question — do drag-and-drop agent builders make sense for anything a developer would ship, or are they how the rest of the org finally gets to build the agents we’ve been talking about all year? — coze.com — github.com/coze-dev/coze-studio

Hermes Agent

Nous Research’s open-source agent that lives on your own box rather than in an IDE — MIT-licensed, self-hosted (a $5 VPS will do), reachable from Telegram, Discord, Slack, WhatsApp, Signal or the CLI through one gateway. The hook is that it’s built to accumulate: persistent cross-session memory, and when it cracks a hard problem it writes itself a reusable skill document, so the agent a month in genuinely isn’t the agent you installed. 64k+ GitHub stars since its February launch and a reported migration wave off OpenClaw. It’s the “agent as a resident” model versus the per-session harnesses this room mostly runs. Open question — do self-written skills compound into real capability, or does the skill pile rot the same way the gstack thread said bolted-on setups always do? — hermes-agent.nousresearch.com — github.com/NousResearch/hermes-agent

Local pipeline for triaging emails with Hermes

The concrete build off the Hermes chat — point it at a dedicated inbox over plain IMAP/SMTP (Python stdlib, no extra services), have it poll for unseen mail, classify each message into buckets (bug, billing, lead, spam), apply labels, and ping you a summary on Telegram via its natural-language cron. Fully local end to end if you swap the cloud LLM for Ollama, so nothing sensitive leaves the network. The sharp edges the room noted — use a throwaway account (the agent gets full inbox access), allowlist senders, and keep triage read-only at first: an email is untrusted input, and a hidden instruction in a “lead” can hijack the agent. Open question — is auto-drafted triage the first agent chore you’d actually trust unattended, and where’s your line between labelling, drafting, and letting it hit send? — Email gateway — Hermes docs

Routines in Claude Code

Anthropic’s new automation layer — configure a job once and it runs on Claude Code’s web infrastructure with no laptop open: on a schedule (“every night at 2am, pull the top bug from Linear, attempt a fix, open a draft PR”), via an HTTP POST to a per-routine authenticated endpoint, or off GitHub webhook events. Routines get access to your repos and connectors, and each PR keeps one session that receives ongoing updates. Pro gets 5 a day, Max 15, Team/Enterprise 25 — set them up at claude.ai/code or with /schedule in the CLI. It’s the same “agents you don’t watch” thread as today’s background subagents, now with a cron. Open question — what’s the first routine you’d actually trust to run at 2am against your repo, and who reviews the draft PRs waiting in the morning? — Introducing Routines — Claude blog

Mixture of experts models

An architecture explainer for the table — MoE swaps the single feed-forward block in each transformer layer for a bank of parallel “experts” plus a router that picks the top-k per token, so the model carries a huge total parameter count but only activates a sliver of it on any given token. That’s the trick behind big-but-cheap frontier models: near-dense-giant capacity at a fraction of the per-token compute, with experts observably specialising by domain. The costs — all those parameters still have to sit in memory, and the router needs load-balancing losses so one hot expert doesn’t bottleneck a GPU. A tidy rhyme with the day’s agent threads: routing work to the cheapest competent specialist, whether that’s a haiku subagent or an expert MLP. Open question — as an API user who never sees the routing, does MoE actually matter to you, or is it exactly the kind of silent internal swap the Fable-nerf thread says we should worry about? — Sebastian Raschka — MoE FAQ — A Visual Guide to Mixture of Experts

ChatJimmy

The speed demo that made the table blink — ChatJimmy is Taalas’ public showcase for Llama 3.1 8B etched directly into a custom ASIC: weights live on the chip, no memory fetch, no software layer, and the result is ~17,000 tokens per second per user against the ~1,800 of a top GPU stack — instantaneous for all practical purposes, with claims of 20× cheaper to build and 10× less power. The catch is the model itself: aggressively quantised, small, limited context, so it’s a proof of “the model is the computer,” not a ChatGPT rival — a mid-size reasoning chip is promised next. Neat foil to the MoE chat above: one architecture routes around compute, the other just deletes it. Open question — if frontier-quality models eventually ship as $30M silicon, what does that do to agent loops built around waiting for tokens — and to the API pricing thread this room keeps circling? — chatjimmy.ai — Taalas’ ChatJimmy — OfficeChai

Alibaba bans Claude Code

The geopolitics slot — Alibaba is banning employees from Claude Code as of July 10, pushing them onto its own Qoder instead, after developers alleged the tool inspects its computing environment (time zones, proxy settings) to spot Chinese users — some called it a backdoor. An Anthropic employee’s response: it was a March experiment to catch account resellers and block model distillation, and it’s being removed. The backstory is the escalation ladder — in June Anthropic accused Alibaba’s Qwen lab of the “largest known” capability-extraction campaign against Claude: ~25,000 fraudulent accounts, 28M+ interactions. Sits right next to the Fable-nerf thread from earlier: the harness on your machine is also a vendor’s enforcement surface. Open question — is environment-sniffing a legitimate anti-distillation defence or exactly the kind of quiet telemetry that justifies the ban, and would your org’s security team pass Claude Code under the same scrutiny? — Alibaba bans Claude Code — American Bazaar

Feishu — Lark

ByteDance’s third appearance of the day, after Seedance and Coze — Feishu is their all-in-one workplace suite, sold internationally as Lark: Slack, Zoom, Google Docs and Notion collapsed into one app, with auto-translated meetings, automatic minutes, OKRs and approval workflows built in. The two brands run separately with data split between Beijing (Feishu) and Singapore (Lark), which is exactly the kind of detail that matters given the Alibaba-ban thread above — trust in tooling now cuts both directions across the same border. For the agent-minded, it’s also an obvious gateway surface: one platform holding chat, docs, calendar and approvals is a lot of context for a bot to sit inside. Open question — would this room ever swap its Slack-plus-five-tools stack for one super-app, and does where the data lives decide it before the features do? — larksuite.com — Lark — Wikipedia

OpenRouter

The plumbing under half the tools discussed today — OpenRouter is a unified, OpenAI-compatible API over hundreds of models from every major lab: one key, one endpoint, automatic fallbacks when a provider degrades, and pass-through pricing with a live public leaderboard of what the ecosystem is actually running. That leaderboard has become the de-facto scoreboard for model adoption — Hermes Agent shows up there as an app — and for BYO-model harnesses like Cline it’s the easiest way to swap brains without touching code. It also cuts straight into the day’s silent-nerf thread: when you route through a middleman, you’re trusting their provider choices too. Open question — is one aggregator key the sane default for agent projects, or does putting a router between you and the model just add another place for quality to quietly change? — openrouter.ai

x402 proxy

The money rail under all the agent talk — x402 revives the dormant HTTP 402 “Payment Required” status into a real handshake: hit a paid endpoint, get a 402 with an amount and a wallet address, your wallet signs a gasless USDC transfer, you retry with an X-PAYMENT header, and you’re through in under two seconds for fractions of a cent. The proxy is the ergonomic bit — an MCP server (or middleware) that catches the 402, signs from the agent’s wallet and retries transparently, so the agent pays per call with no API key, no signup, no human. Coinbase open-sourced it and donated it to a Linux Foundation body co-run with Cloudflare; ~69k active agents already. It reframes today’s threads hard — the ghealthorg funnel paying agents in USDC, or an agent renting Seedance by the shot, stop being hypothetical. Open question — do we actually want agents holding wallets and spending autonomously per request, and what’s the blast radius the first time a prompt-injected agent goes on a shopping spree? — x402.org — github.com/coinbase/x402

Alipay AI Wallet — Token Pay

The incumbent-money counterpart to the x402 chat — in May 2026 Ant Group shipped two pieces of agentic-payment plumbing: AI Wallet, a consumer control layer in the Alipay app where you set limits and authorise what your agents spend before, during and after a purchase, and Token Pay, a B2B rail for model providers to handle subscriptions, token top-ups and microtransactions between platforms and merchants. CEO Cyril Han’s framing — “agents execute payments, tokens carry value” — splits authorisation from settlement, the same seam x402 draws. Where x402 is crypto-native and permissionless, this is a walled billion-user wallet with a “trust protocol” bolted on — and a callback to the Alibaba/Claude Code ban, another front in the same East–West stack war. Open question — when agents start spending for us, do we want the crypto rail or the mega-wallet holding the leash, and which does your threat model actually trust less? — Alipay AI Wallet & Token Pay — Stellagent

AP2 — Agent Payments Protocol

The trust layer the payment thread was missing — Google’s AP2 doesn’t move money itself, it proves intent: cryptographically-signed Verifiable Digital Credentials called “mandates” that pin down what a user actually authorised an agent to buy, for how much, and when. A Checkout Mandate captures the purchase, a Payment Mandate the spend, and both work whether the human is present or not — a tamper-evident audit trail for “did the user really okay this?” It’s positioned as an extension to Google’s A2A protocol and, tellingly, ships x402 integration in its samples — so it’s the accountability wrapper over rails like x402 and the Alipay wallet, not a competitor. Standardisation is being pushed through the FIDO Alliance. Open question — is a signed mandate enough to make an agent’s spend defensible when it goes wrong, and who eats the loss when a prompt-injected agent produces a perfectly valid-looking mandate? — ap2-protocol.org

Worldcoin

The proof-of-personhood turn on the payments thread — Sam Altman’s Worldcoin scans your iris with a chrome “Orb” to mint a World ID, a unique credential that proves you’re a distinct human without revealing who you are, paired with a crypto token dropped to verified sign-ups. It’s been dogged from the start by privacy pushback and regulator bans over harvesting the world’s most sensitive biometric to bootstrap a token. But it reads differently against today’s agent-economy arc: once x402 wallets, AP2 mandates and mega-wallets let software transact freely, “is there a real human behind this?” becomes the scarce signal — and an iris might be how you gate agents out. Open question — in a web swamped with agents, is biometric proof-of-personhood the necessary bot-filter, or a cure worse than the disease we’d regret handing one company? — Worldcoin explained — Fortune — world.org

Idiocracy

The half-joking coda to the whole morning — Mike Judge’s 2006 satire where an average guy wakes 500 years later to find civilisation dumbed itself into collapse: crops watered with sports drink, a machine that just says “please wait” while everything rots. It came up as the nervous punchline to the day’s threads — agents that auto-commit unwatched, biometric orbs, a health funnel paying crypto to route you to stem-cell clinics. The worry the table circled: if we hand judgement to the machines and stop verifying, is the failure mode a dramatic robot uprising or just a slow, comfortable Idiocracy where nobody remembers how anything works. Open question — is “keep a human who actually understands the system in the loop” the real lesson under all today’s automation, and are we already outsourcing the understanding, not just the work? — Idiocracy — Wikipedia

Universal basic income

Where the automation-anxiety thread lands — UBI is the idea of paying every person a regular, unconditional sum with no means test and no work requirement, floated as the cushion if agents and models genuinely eat the jobs. It’s a direct callback to the Engels’ Pause chat from AA04: if productivity keeps climbing while wages don’t, UBI is one answer to who catches the people the curve leaves behind. The room split the usual way — a floor that frees people to retrain and build versus a bill nobody’s shown how to fund without inflating it away, plus the darker read that it’s severance for being made economically redundant. Sharpens against today’s payment rails: the same infrastructure that lets agents earn USDC could just as easily pipe a stipend to humans. Open question — if this room’s own tools are part of what makes work optional, is UBI the pressure valve or just a way to keep everyone quiet in the Idiocracy? — Universal basic income — Wikipedia

News since last assembly

Floor: 2026-06-27 (aa07). Generated 2026-07-04.

New Claude Code commands & features

• Claude Sonnet 5 is the default model (v2.1.197, 2026-06-30) — native 1M-token context; promotional pricing $2/$10 per Mtok through Aug 31 — release
• Subagents run in the background by default (v2.1.198, 2026-07-01) — Claude keeps working and is notified when they finish (Notification hook: agent_needs_input / agent_completed); background agents from claude agents now auto-commit, push, and open a draft PR; set run_in_background: false to force foreground — release
• /dataviz skill (v2.1.198, 2026-07-01) — chart and dashboard design guidance with a runnable colour-palette validator — release
• Explore agent inherits your model (v2.1.198, 2026-07-01) — the built-in Explore agent now uses the session model (capped at opus) instead of always running on haiku; the /agents wizard was removed in favour of editing .claude/agents/ directly — release
• Stacked slash-skill invocations (v2.1.199, 2026-07-02) — /skill-a /skill-b do XYZ now loads all leading skills (up to 5), not just the first — release
• Default permission mode → Manual (v2.1.200, 2026-07-03) — new default across CLI, VS Code and JetBrains; AskUserQuestion no longer auto-continues; screen-reader output substantially improved — release
• Org default models (v2.1.196, 2026-06-29) — admins set an org-wide default model (shows as “Org default” / “Role default” in /model); MCP servers in untrusted workspaces now require explicit approval — release

Codex

• [2026-07-01] Quiet stretch — only 0.143.0-alpha builds and a logging bug-fix (0.142.5); no new commands or features landed since the floor — releases

Adjacent tools

• [2026-06-29] Cursor for iOS ships in public beta on all paid plans — launch always-on cloud agents in isolated VMs, Remote-Control local agents, steer by voice and slash command, review diffs/screenshots from your phone; 75% off Composer 2.5 runs through July 5 — Cursor blog
• [2026-07] Cursor Team Marketplaces expand to Team MCPs — admins configure MCP servers once and distribute them across cloud agents, the IDE and the CLI — Cursor updates
• [2026-07] Claude in Chrome reaches general availability — changelog

Simon says

• [2026-07-03] “Fable’s judgement” — let the model decide when to write tests rather than dictating it; push trivial edits to haiku subagents and keep judgment-heavy work in the main model to save tokens before prices rise — post
• [2026-07-02] Releases llm-coding-agent 0.1a0, a Claude-Code-style coding agent built on his LLM library — post
• [2026-06-28] “Hack Your Summer” — a free coding cohort; the second cohort starts July 13 (student applications due July 8) — post

Research & papers

• [2026-06-30] MIT Q&A on agentic AI — coding agents are the clearest success story, but easy “vibe coding” tempts people to ship unverified code, risking bugs and leaked data — MIT News

Notable posts

• [2026-07] The MCP 2026-07-28 release candidate enters its validation window — the largest revision yet: a stateless protocol core over plain HTTP, an Extensions framework, long-running Tasks, server-rendered MCP Apps, and hardened OAuth/OIDC authorization; final spec ships July 28 — Model Context Protocol

Topics worth a 5-min slot

1. Agents you don’t watch — subagents now run in the background and auto-open draft PRs, and Cursor puts cloud agents on your phone. Where’s the line between “unlock” and unsupervised token-burn, and who reviews the auto-pushed work?
2. Sonnet 5 as the new default — a 1M-context model at promo pricing became Claude Code’s default overnight. Does it change how you delegate, and does the promo end-date change your token math?
3. MCP goes stateless — the July 28 spec drops the stateful core for plain HTTP plus Extensions, Tasks and MCP Apps. What breaks, and what does a stateless MCP unlock for the servers this room runs?