ClaudeDevs disclosed it directly: Claude Tag already merges 65% of the Claude Code team's PRs, including most of the code that built Claude Tag itself. The same week, Justin Poehnelt went public that Google fired him for shipping a Workspace CLI that hit number one on Hacker News — two days before Google Cloud Next announced an official Workspace CLI. One organization let an agent take over its own build loop. The other fired the person who showed the gap. That gap is the story.
What's underneath it is a split in how incumbents and builders read the threat. Anthropic is embedding Claude inside Slack with memory, tools, and scheduling — the model stops being a destination and becomes the coordination layer. Google is protecting the coordination layer it already owns. Gergely Orosz called the Claude Tag move correctly: this is a tooling-distribution play, not a model-quality play. The companies treating agents as a product to defend against will keep firing the Poehnelts. The companies letting agents merge their own PRs will compound the lead. Those two strategies don't end in the same place.
Top developments
Anthropic — launched Claude Tag: Claude joins Slack as a persistent team member tied to channels (not users), with memory, tools and autonomous task scheduling; ClaudeDevs disclosed that Claude Tag already merges 65% of the Claude Code product team's PRs internally, including most of the code that built Claude Tag itself. Karpathy framed it as the third major redesign of LLM UIUX after website and downloadable app — the model stops being a thing you go to and becomes a persistent async teammate sitting inside the org's coordination layer. x x
Krea AI — released open weights for Krea 2 Raw and Krea 2 Turbo under a custom license: an undistilled mid-training checkpoint meant to be fine-tuned, plus a fast distilled variant pitched at enterprise-grade image generation in roughly two seconds. The open-weights frontier keeps marching down the stack from text to image; "fine-tune your own image model" just stopped being a frontier-lab privilege. x x
Justin Poehnelt (ex-Google) — went public that Google fired him for building the Google Workspace CLI that hit #1 on Hacker News, thousands of GitHub stars and thousands of users in days; he points at internal Workspace leaders afraid of agent-led disruption, two days before Google Cloud Next announced an official Workspace CLI. Distribution incumbents are now firing the people most likely to expose how exposed they are — and the story going viral the same day Anthropic ships Claude Tag is its own commentary on which side moves faster. x
Anthropic — opened "Claude for Startups": free Claude API credits, top-tier rate limits with no production throttling, and early access to model releases for early-stage VC-backed founders backed by Anthropic partner VCs. Coming the same day as Claude Tag, the developer-economy plumbing is the actual play — Anthropic is buying the next cohort of agent-builders before they ever evaluate an alternative. x
Notable discussions
Frontier-model release cadence is slipping — synthwavedd's scoop: GPT-5.6 pushed to mid-July, DeepMind unhappy with Gemini 3.5 Pro and pulling its launch, Claude Sonnet 5 quietly shipped as enterprise stop-gap while Mythos/Fable 5 stay export-stuck; Arthur Mensch confirmed on CNBC that OpenAI and Anthropic are calling Mistral asking for compute; Snowflake's CEO publicly benchmarked GLM-5.2 against Opus-4.7 on dbt-bench and signaled it's "good enough" to roll out to customers; cryptopunk7213 read it as Anthropic obviously about to distill its flagship into something cheaper. Three months ago the frontier story was who would ship next; this week it's who can keep the lights on, with capacity, export controls and open-weight pressure all hitting the schedule at once. x x x x
Karpathy — calls Claude Tag the third major redesign of LLM UIUX: paradigm one was the LLM as a website, paradigm two was the downloadable app, paradigm three is a persistent async entity with org-wide tools, memory and context working alongside humans — and argues you have to spend time with it before the shift lands. x
The_Prophet — reads Claude Tag as marketing-as-collaboration, structure-as-labor-absorption: once an agent is inside Slack with permissions and memory, the coordination-heavy middle of the org — following up, summarizing, status drift, prep, PM glue — gets quietly compressed, high-agency operators absorb more territory, and the headcount math changes without anyone announcing layoffs. x
Gergely Orosz — argues Anthropic just stopped competing on "best model" and started competing on tooling distribution: Claude Tag + Cowork + Code + Managed Agents is an integration-across-workflows play, and the CTO move now is a Slack integration that lets you switch models any time to avoid being locked in. x
Linus Torvalds — gets visibly angry at the Open Source Summit when the room repeats the "99% of code is AI-written" claim, the first prominent rejection of the headline number that's now load-bearing for the loop-engineering frame in this newsletter and elsewhere. x
Richard Sutton (via phosphenq) — uses his MIT talk to argue today's chatbot-first approach is the wrong path to general intelligence: the next paradigm is self-taught agents learning from play and abstraction, "a path towards intelligence that isn't limited by human abilities" — the Turing winner saying the dominant lab playbook is a dead end. x
Demis Hassabis (via ihtesham2005) — confirms every frontier lab is actively running recursive self-improvement loops in coding and math where verification is fast, says the unsolved problem is biology/chemistry where the loop closes in weeks not seconds, and says removing humans from the loop entirely keeps him up at night — the most explicit on-record acknowledgement from a frontier-lab CEO that RSI is the active program, not a future risk. x
Other news
Models & releases
Anthropic Sonnet 5 — shipping to select enterprise customers under an Early Access Program as a stop-gap while Fable 5 / Mythos 5 stay export-blocked x
OpenAI Bidi — new bidirectional voice mode in active prep for ChatGPT, potentially launching this week x
Meta Mythos-level model — reportedly trained, Q3/Q4 release window x
Microsoft bitnet.cpp — 100B-parameter models running on consumer hardware x
Tencent Agent Memory — open-source 4-tier semantic memory pyramid, 61% token reduction, 76% on PersonaMem x
Gemma 4 — hits 300 tok/s aggregate throughput on a single DGX x
Baidu Unlimited-OCR — processes entire documents without chunking x
Mistral OCR 4 — adds bounding boxes for form-filling accuracy x
HyperQuant — lattice-based compression for LLM key-value cache x
Devtools & coding agents
Linzumi — Codex but multiplayer, from Sean Grove (ex-OpenAI sycophancy team), YC S26 x
Aside — YC's AI browser launches with vertical tabs, Liquid Glass and agentic capabilities x
LiteLLM — moving to Rust: sub-1ms overhead, sub-100MB binary, same Python SDK x
Vercel tsgo — frontend monorepo type checks drop from ~2 minutes to 14 seconds, 7x faster x
Vite 8.1 — experimental bundle mode accelerates large-app dev servers x
Hermes /learn — point Hermes at a codebase or set of docs and it crystallizes a reusable skill, not just memory x
Momentic — QA agent ingests every Linear ticket, Notion PRD and PR; 73% PR merge rate, customers include Notion, Xero, Webflow, Retool, Runway x
Executor — YC S26 open-source MCP gateway for agent-service connections x
Apple Foundation Models — framework now opens to cloud providers, not just Apple silicon x
Apple containerization — Docker Desktop now optional and free; native Linux containers on macOS x
MongoDB Cursor plugin — bundles MCP server with pre-built agent skills x
Modal Auto Endpoints — true ownership of AI inference infrastructure x
Mastra — agents now support structured task lists for organized workflows x
Warp — adds GLM-5.2 support with flexible inference options x
Firecrawl — becomes official Grok plugin for web search and scraping x
Mercury Command — handles mission-critical workflows with human review safeguards x
Infrastructure & platforms
Industry & policy
Anthropic-Persona ID verification — power-user concerns surface again as KYC rollout continues (already in tail 06-23) x
Congress + Lutnick — four members of Congress demand explanation of Howard Lutnick's Fable export ban by June 26 x
Anthropic + Chad Jones — hires Stanford growth-theory economist Chad Jones x
Slash / $80K token burn — fintech rolled back its AI coding push after one employee burned $80K in tokens in week one x
Brilliant Labs Halo — AR glasses with full-color display, on-device AI, 14-hour battery x
Funding & deals
Continuing threads
Sakana Fugu (cont.) — Hesamation accuses Sakana of rebranding orchestration as "single foundation model" and farming attention via misleading creator framing x
Anthropic cofounder / singularity 2028 (cont.) — Hesamation surfaces the line: "Claude 10 would build Claude 11 without any researchers involved" x