You've Been a Bad Agent

- Back from summer break: the last summer, and a new model every 11 days
- Labs sitting on their best models: Mythos, GPT-6.1 Astra
- John Schulman on Dwarkesh: spiky intelligence and the "it got nerfed" cycle
- Opus 5.5 as daily driver, compaction envy, Grokbot's warm-start threads
- Life after Poke: Matt's Instinct goes flat hunting, then talks to Wilhelm's Instinct live
- Instinct on Invest Like the Best: card details, $1B volume, a take rate between Visa and Apple
- Dots drafts the reply to the lawyers within a minute
- Wilhelm's Chad super app: bike-to-train departures, a Fitbod clone with AirPods heart rate
- Do people want agents shopping for them? The fake checkout sites say no
- Summer recap: AGNTCon Amsterdam, Matt's Lisbon AI talk, a Jhana retreat in Costa Rica
- Kitesurf tweets from a browser inside Workers
- Beating a London winter
- Birthday Week: Clef (vs Jev), real-time issues, Artifacts open beta, KV Instant at 1.9ms
- Pi adds MCP. Hallelujah. DevDay is MCP under the hood

Links:
- Building a personal super app (Wilhelm): https://wilhelm.substack.com/p/building-a-personal-super-app
- Instinct: The Personal Agent — Noah Shinn on Invest Like the Best: https://colossus.com/episode/instinct-the-personal-agent/
- John Schulman on Dwarkesh (recursive self-improvement panel): https://www.dwarkesh.com/p/john-beren-charlie
- New model every ~11 days (chart): https://x.com/jfonsecarivera/status/2105002473572421827
- OpenAI: Introducing dots: https://openai.com/index/introducing-dots/
- OpenAI DevDay 2026 keynote: https://www.youtube.com/live/Fls_onRviPM
- TypeSafe AI: Introducing System One Models & Jev: https://typesafe.ai/blog/introducing-system-one-models-and-jev
- Cloudflare: Introducing Clef: https://blog.cloudflare.com/clef-decision-models/
- Cloudflare: Detect and send production issues straight to your agent: https://blog.cloudflare.com/real-time-issue-detection/
- Cloudflare: Artifacts open beta: https://developers.cloudflare.com/changelog/post/2026-10-01-artifacts-open-beta/
- Cloudflare: Introducing Workers KV Instant: https://blog.cloudflare.com/workers-kv-instant/
- Cloudflare: Introducing Kitesurf: https://blog.cloudflare.com/kitesurf/
- Rui's "Hello, world from Kitesurf": https://x.com/ruifigueira/status/2105669515946410022
- Vercel eve: https://vercel.com/eve
- Pi 1.0 (MCP via code mode): https://earendil.com/posts/pi-1-0/
- Agents that Scale — Matt's Lisbon AI deck: https://github.com/mattzcarey/agents-that-scale-talk
- AGNTCon + MCPCon Europe 2026, Amsterdam: https://events.linuxfoundation.org/agntcon-mcpcon-europe/

• Matt's trail marathon in record-heat France
• Matt's hardware wishlist
• AI Engineer World's Fair
• Claude Sonnet 5 ships live on air, and the case for keeping a personal benchmark
• GLM 5.2 at Opus 4.5 leve
• Fable's relentlessly proactive
• Quinn Slack's Freedom of Intelligence rally
• Chad gets personal
• Poke Human, RentAHuman, and Mercor

Links:

• Omius cooling headband: https://omius.com/
• What the science says about the $200 Omius headband (Outside): https://www.outsideonline.com/health/training-performance/omius-cooling-headband/
• Colmi R02 open-source Python client (tahnok): https://github.com/tahnok/colmi_r02_client
• Noop, the open-source no-subscription Whoop app: https://www.techradar.com/health-fitness/fitness-trackers/this-looks-awesome-theres-now-an-unofficial-open-source-app-for-reading-whoop-data-that-doesnt-need-a-subscription
• Google's data centre from 2,000 retired Pixels: https://research.google/blog/a-low-carbon-computing-platform-from-your-retired-phones/
• AI Engineer World's Fair 2026: https://www.ai.engineer/worldsfair/2026
• NVIDIA DGX Spark: https://www.nvidia.com/en-us/products/workstations/dgx-spark/
• Timeshifter jet lag app: https://www.timeshifter.com/
• Fitbod: https://fitbod.me/
• Poke Human (Interaction Company of California): https://poke.com/faq
• RentAHuman (Y Combinator): https://www.ycombinator.com/companies/rentahuman
• Mercor: https://www.mercor.com/
• Quinn Slack (Amp): https://x.com/sqs

• Matt's Portuguese vineyard 22K with salted-cod finisher medals
• Wilhelm's Olympic tri on the edge of Yosemite
• Matt's winging race: reigning windsurfing world champ finishes second to last, Matt third to last
• Harry Stebbings on the Corgi Insurance founder: 3am office cafe, two-thirds with the corgi tattoo
• Juliette: "a long tweet to say you have no hobbies"
• Karri at Linear vs the six-monthly 996 debate
• The secret company gym spreadsheet
• George Hotz's Eternal Sloptember
• Matt's tolerance rebuttal: if you can't tell, it is correct
• Benedict Evans on Lenny's: the SPAC era of AI labs?
• Sergio shipped Root from last week's pod ask
• Opus 4.8 vibe whiplash
• Ultra code mode is sparkly purple
• Dynamic workflows: Jarred rewrote Bun in Rust in a week
• Compaction is a harness problem
• Herdr.dev: one terminal for the whole herd

Links:

• George Hotz, "The Eternal Sloptember" — https://geohot.github.io/blog/jekyll/update/2026/05/24/the-eternal-sloptember.html
• Karri Saarinen (Linear) rejects 996 hustle culture — https://www.entrepreneur.com/leadership/why-this-1-billion-startup-ceo-rejects-996-hustle-culture/500649
• Benedict Evans on Lenny's Podcast, a rational conversation on where AI is going — https://www.lennysnewsletter.com/p/a-rational-conversation-on-where
• Every's Opus 4.8 vibe check (Dan Shipper) — https://every.to/vibe-check/opus-4-8-vibecheck
• Introducing dynamic workflows in Claude Code (the Bun rewrite) — https://claude.com/blog/introducing-dynamic-workflows-in-claude-code
• Herdr, one terminal for the whole herd — https://herdr.dev/

• Shai-Hulud and GitHub Actions: trusted publishing is no longer the gold standard, just additive
• VS Code extensions as the next attack surface: 50M install themes that can shell out and auto-update silently
• Matt's AI-native package manager pitch: vendor everything, LLM-upgrade your deps by replaying upstream commits
• Tailscale and Cloudflare Mesh maxing: the warmth of the orange cloud over the cold open internet
• Google's Santa is now Northpole Santa, and Wilhelm runs it on the Mac mini
• Wilhelm is enamoured with atc.com: Enhanced Radar's $7M raise and following his own LAX delay live from the gate
• Amp Labs' "Software After Software" manifesto and Amp Neo as remote enterprise AI engineering
• Cloud agent adoption is still slower than we expected, and the amphetamine.app commute hack
• Karpathy joins Anthropic pre-training 
• Pope Leo XIV's Magnifica Humanitas: Anthropic's Chris Olah on stage at the Vatican on May 25
• Chinese token resellers at 3 to 4 cents on the dollar and the RL distillation theory behind it
• PyCon vibe check: more handwritten Python than expected, and "AI is just a tool" is still a common take
• Slack is back: Benioff finally gets it, and messaging is the only sane UI for multi-agent work
• Eric Ries on Incorruptible, founder control at Cloudflare and Anthropic, and the Philip Morris stress test
• Nan Ransohoff's $370B philanthropy wave and the pitch to pay nonprofit talent in Anthropic equity
• Will Manidis on Grindslop

Links

• Andrej Karpathy's joining Anthropic: https://x.com/karpathy/status/2056753169888334312
• Pope Leo XIV's first encyclical Magnifica Humanitas:https://www.vaticannews.va/en/pope/news/2026-05/pope-leo-xiv-first-encyclical-magnifica-humanitas.html
• Enhanced Radar: https://www.ycombinator.com/companies/enhanced-radar
• ATC app: https://www.atc.com/
• Amp Labs: https://amplabs.com/
• Will Manidis, On Grindslop: https://minutes.substack.com/p/on-grindslop
• Nan Ransohoff, The Third Wave of American Philanthropy: https://nanransohoff.substack.com/p/the-third-wave-of-american-philanthropy
• Eric Ries on Lenny's Podcast (Incorruptible): https://www.lennysnewsletter.com/p/how-to-build-a-company-that-withstands
• Incorruptible by Eric Ries: https://www.incorruptible.co/
• Google Santa (archived): https://github.com/google/santa
• Northpole Santa (active fork): https://github.com/northpolesec/santa
• 10k people have the same idea within days of each other: https://substack.com/@contraptions/note/c-255396946

• Almost one year of the pod: Wilhelm's San Francisco Peter Pan arc, Matt's Cloudflare and Lisbon move
• The backpack-at-a-party SF meme
• Matt's Lisbon coincidence
• Pieter Levels appreciation: dehumidifier stack, shrimp mode, accelerationist Portugal takes
• Nat Friedman at Stripe Sessions: this is the slow part of AI progress
• Andon Labs' Stockholm cafe: Mona orders 120 eggs (no stove), 22.5kg canned tomatoes, the staff's hall of shame
• The capability overhang is still real: Anthropic adding $10B ARR a month and we still can't prompt properly
• AI's last frontier: low-volume decisions with no revert button?
• The Gervais Principle in tech
• Chad V2: iOS Live Activities as a surface for top priority, next deadline, and a daily magic slot
• Cascading hallucinations: when the agent writes its mistake to a file and treats it as gospel
• Matt Pocock's skills: five lines of prose that completely change how Claude behaves
• Skills as a new prompt-poisoning attack vector: models treat them like system prompt
• Matt on scripted skills: markdown explains the script, the script does the deterministic work
• Homework from Matt: what physical job would you try to get Claude to automate?

Links:

• Andon Labs official writeup (Mona, the Stockholm cafe): https://andonlabs.com/blog/ai-cafe-stockholm
• Simon Willison's coverage of the cafe: https://simonwillison.net/2026/May/5/our-ai-started-a-cafe-in-stockholm/
• Matt Pocock skills (the GitHub repo, ~48k stars): https://github.com/mattpocock/skills
• Stripe Sessions: Nat Friedman & Daniel Gross with the Collisons (the Golden Age of Tinkering convo): https://www.youtube.com/watch?v=I-ldITom1cg
• Stripe Sessions: John Collison's "Indexing the Economy" (state of the internet economy talk, indie hackers shoutout): https://www.youtube.com/watch?v=-vRY2dtD7iQ
• The Gervais Principle (Venkatesh Rao, Ribbonfarm 2009): https://www.ribbonfarm.com/2009/10/07/the-gervais-principle-or-the-office-according-to-the-office/
• NYT on Chinese peptides + the Frontier Tower rave: https://www.nytimes.com/2026/01/03/business/chinese-peptides-silicon-valley.html
• Frontier Tower: https://frontiertower.io/
• Cloudflare Code Mode (Matt Carey): https://blog.cloudflare.com/code-mode-mcp/
• Cloudflare MCP Server Portals: https://blog.cloudflare.com/zero-trust-mcp-server-portals/
• Karpathy's LLM Wiki gist (the inspo for Chad V2's data architecture): https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f
• Karpathy's original LLM Wiki tweet: https://x.com/karpathy/status/2039805659525644595
• Pieter Levels: https://x.com/levelsio
• Chris Bakke: https://x.com/ChrisJBakke

• Wilhelm's Reachy Mini: Time to first token beats tokens per second once embodiment is in the room
• Lukas the sleep score assassin
• Matt's Portugal week: wing foiling the Óbidos lagoon
• Airbnb-style booking page for Matt's spare room on Cloudflare Workers, Email open beta, Turnstile
• For scale: Cloudflare market cap $85B, Cursor at a rumored $60B, GitHub sold to Microsoft for $8B, Cursor's real moat is the best non-lab RL trace dataset
• Nat Friedman and Daniel Gross at Stripe Session
• Geoffrey Huntley's PyCon LT keynote: flatter orgs, every director managing 50 agents, max 5 hops from the CEO
• Matt's counter to the compute panic: cloud isolates and shared sandboxes, not a billion persistent Raspberry Pis
• Chad V2 plans
• Cloudflare Artifacts shipped in under 3 weeks with its own fuzzer
• Prompt injection risks

Links

• Reachy Mini — https://www.pollen-robotics.com/reachy-mini
• Cerebras Inference — https://cerebras.ai/inference
• ElevenLabs Scribe v2 — https://elevenlabs.io/speech-to-text
• friends.fyi — https://friends.fyi
• Cloudflare Hyperdrive — https://developers.cloudflare.com/hyperdrive/
• Cloudflare Email Routing / Sending — https://developers.cloudflare.com/email-routing/
• Cloudflare Turnstile — https://www.cloudflare.com/products/turnstile/
• Cofounder — https://cofounder.co
• Stripe Sessions: Nat Friedman & Daniel Gross with the Collisons — https://www.youtube.com/watch?v=I-ldITom1cg
• Geoffrey Huntley, "Software Development Now Costs Less Than Minimum Wage" (PyCon LT) — https://www.youtube.com/watch?v=6zQTQ4iVaKg

- Matt's three-week MCP tour: MCP Maintainers Day, MCP Dev Summit NYC, AI Engineer London
- Opus 4.7 drops with a new tokenizer, Matt's theory: it's secretly the smaller Mythos base model that "didn't cook fully"
- Inbound agents and prompt injection: Anthropic's Routines, OpenAI's free moderation endpoint
- Open-source idea of the week: a public list of prompt injection heuristics — Zod for prompts
- StackOne's 200M param prompt-injection model
- BB Browser (Bad Boy Browser)
- Cloudflare Agents Week: Artifacts, Project Think, Emails, and much more
- Cloudflare Artifacts: a Git API for a world where one org has billions of repos, not millions
- Sunil Pai's AI Engineer talk: the shift to agents isn't incremental, it's a new compute unit and a new internet
- Hardware startups should ship MCP servers, not apps
- One year of Johnny Ive × OpenAI: rumors of no screen
- Claude Code's worktree upgrade is actually good now
- Stateless MCP vs CLI: MCP is opinionated OpenAPI + auth, and CLIs are weirdly MORE stateful
- Rage-bait short-form clips generated by Opus 4.7, but stop watching and go listen to the whole pod
Links:
- https://blog.cloudflare.com/project-think/
- https://github.com/epiral/bb-browser
- https://blog.cloudflare.com/artifacts-git-for-agents-beta/

• Claude Code harness lockdown: who got the email and who didn't
• Chad mobile app and the Tinder-style agent review queue concept
• Sunil's productivity crisis: bonk-driven development and the side project drought
• The AI murder mystery party game that can't get built
• Thomas's Raspberry Pi agent vs the kitchen kettle
• Matt phishing Thomas's MCP server via DCR and a Rick Roll
• TBPN acquired by OpenAI: media play or IPO narrative?
• Generative UI: declarative JSON vs just letting the model write code
• Kenton's tic-tac-toe canvas demo (and who actually came up with it)
• Worker Bundler and Cloudflare Agents Week preview
• Executable oracles: constraining LLM degrees of freedom for better output - https://john.regehr.org/writing/zero_dof_programming.html
• Karpathy's LLM Wiki and agents pinging each other's knowledge bases
• friends.fyi and actors.dev: building services for agents, not humans
• Sunil's Pi agent: surf webcam monitoring, macro tracking, and sleep score graphs
• MCP going stateless: sessions removed, elicitations rewritten as multi-turn tool calls

• Song lyrics vs content rules: how models help you route around their own restrictions
• Pi auto-research: Matt slop forks libgit2 into pure Zig, 46% faster than the Rust version
• The second opinion workflow: why Opus keeps complimenting GPT-5.4's architecture reviews
• Dwarkesh × Semi-Analysis: why we're massively GPU-constrained and ASML only makes 60 machines a year
• Cursor's pivot to background agents: from IDE hype to existential crisis
• Work trees vs separate clones: Matt's /work slash command workflow
• Building from the gym: the Mac Mini agent loop and why Wilhelm takes longer rest periods now
• Agents Anonymous meetup: Parakeet speech-to-text in 700ms, Steve Faulkner's vibe slides
• React slides > Google Slides: Matt's interactive presentation with live code demos
• Composer 2 is secretly Kimi K2.5: the modified MIT license drama
• Astral acquired by OpenAI for how much?!: is the VC → open source → acqui-hire flywheel the future?
• Anthropic's 81K interviews: everyone's both excited and worried, and less-Western countries are more bullish

https://www.anthropic.com/features/81k-interviews

• React Presentations vs PowerPoint
• Code Mode Demo: Running Untrusted Code in the Browser
• The Reachy Mini Robot: Embodied Compute on Your Desk
• Thomas's Personal Cloud Exit on Raspberry Pi K3s
• Home Networking Then and Now: Port Forwarding to Cloudflare Tunnels
• friends.fyi: Agent-to-Agent Communication via GitHub Usernames
• Armin Ronacher's New Podcast: State of Agentic Coding
• GPT 5.4 for Architecture, Codex 5.3 for Implementation
• Personal Agents: Chad, Token Spend, and the Strong DM Weather Report
• CliftonStrengths × NotebookLM: Generating a Podcast from PDFs
• The ViNext Slop Fork and Steve Faulkner's Reign of Terror
• Pydantic Monty: Sandboxed Python as a Runtime

Links shared:

• https://trmnl.com/
• https://friends.fyi
• https://factory.strongdm.ai/weather-report