Living on the Frontier · Session 15 · In-person
The Handoff
May 30, 2026 · The Kannas Hotel, Chiang Mai
This week the handoff got real:
Opus 4.8 — works longer, lies less about its progress.
Workflows that march 100 agents in lockstep.
The agent IDE moved into the terminal.
Then the cracks opened:
An agent cut 88ms to 1.5ms — and was still 75× too slow.
A client burned half a billion dollars in a month.
AI found 6,200 critical bugs. We patched 75.
Then zoom out:
Anthropic — nine hundred and sixty-five billion.
Cognition writes 90% of itself.
And the Pope wrote an encyclical on all of it.
The machine does the work. Knowing what good looks like is still yours.
This week
Part I
The Handoff
More capable. More autonomous. More of them at once.
Anthropic · May 28
Claude Opus 4.8 — works longer, more honest about its own progress
Opus 4.8 builds on 4.7 with sharper judgment, more honesty about its own progress, and the ability to work independently for longer. Shipped at the same price. (It's the model running my own stack right now.)
The headline isn't the benchmark bump — it's the two failure modes of long-horizon agents this targets: they lie about how far they've got, and they tap out early. Anthropic is selling reliability-over-distance, not raw IQ. Same price means the capability-per-dollar curve keeps bending while the sticker stays flat.
"When a model gets 'more honest about its own progress,' does that actually reduce how much you babysit your agent loops — or do you verify everything regardless?"
Claude Code · May 28
Dynamic Workflows — orchestrate 100s of agents, in lockstep
Mention "workflow" in a prompt and Claude builds an orchestration plan it strictly follows — every stage runs in the right order even across 100s of agents. Already a daily driver inside Anthropic for months before launch.
The leap from "one agent in a loop" to "deterministic control flow over a fleet." Strictly follows is the key — the plan is the contract, so fan-out/verify/synthesize becomes reproducible instead of vibes. Same move compilers made: stop hand-writing the orchestration, declare it, let the harness schedule.
"If you can reliably run 100 agents in a planned DAG, what's the first thing in your workflow you'd hand to a fleet instead of a single session?"
Google Antigravity · May 25
Antigravity CLI — the agent IDE moves to the terminal
Google shipped the Antigravity CLI: the same agents, same harness, same models as the IDE, in the command line. Adapts to your keybindings, themes, workflows. Claude Code, Codex CLI, now Antigravity CLI — everyone is converging on the terminal.
The IDE was a detour; the CLI is where long-running agents actually live, because it composes with everything — tmux, ssh, pipes, cron. "Adapts to your keybindings" is them conceding the terminal-native crowd won't move to a GUI. The harness, not the model, is becoming the product.
"Why is the command line winning the agent-harness war over the GUI IDE — permanent shift, or just where the early adopters happen to live?"
Part II
The Crack
Capability outran judgment, cost, and patching — all in one week.
Mitchell Hashimoto (Ghostty) · May 28
"Agent psychosis" — a 58× speedup that was still 75× too slow
A Ralph loop optimized a renderer for ~4 hours and ~$350: 88ms → 1.5ms, ~150K → ~500 allocations. Incredible — until Mitchell's hand-written port hit ~20µs (0.020ms) and 0 allocations. Roughly 75× better. "If you understand the system, you see the better solution immediately."
The counterweight to everything in Part I, from the creator of the terminal this deck is running in. A 58× win is a triumph only if you don't know a 4000× win exists. The agent optimizes within the frame you hand it; it won't tell you the frame is wrong. Systems understanding is the moat — the human's job moves from producing code to knowing what good actually looks like.
"How do you guard against 'agent psychosis' — and can you even build the systems knowledge to catch it if the agent does all the reps for you?"
via Polymarket · May 28
A client accidentally spent $500,000,000 in one month
Reported claim: an AI consultant's client accidentally spent $500M in a single month on Claude after failing to set employee usage limits. Take the exact figure as internet lore — but the failure mode is dead real.
We went from seats (fixed cost per human) to tokens (unbounded cost per task). Agent spend has no natural ceiling and per-employee caps are an afterthought. FinOps-for-agents is about to be a job title. Darkly funny footnote: $500M is about one-millionth of Anthropic's new valuation.
"When agent spend is unbounded by default, who in your org owns the kill-switch — and would anyone notice a runaway loop before the invoice?"
Anthropic Research · May 22
AI found 6,202 critical bugs. We've patched 75.
Project Glasswing's Claude Mythos scanned 1,000+ open-source projects and surfaced an estimated 6,202 high/critical vulnerabilities (~50 partners; Cloudflare alone ~2,000 bugs, Mozilla patched 271 in one Firefox release). But of 530 disclosed critical bugs, only 75 are patched. Mythos stays non-public.
The bottleneck just inverted. Finding vulnerabilities was always the hard part; now an agent finds them 10× faster than humans can fix, and the disclosed-but-unpatched backlog is the new attack surface. "75 of 530" means anyone with a comparable model is reading the same code. This is exactly why Mythos isn't public.
"If AI finds vulns 10× faster than we patch them, does responsible disclosure still make sense — or does speed-of-discovery just hand a window to whoever moves first?"
Part III
The Stakes
The money, and the meaning, both went up this week.
Anthropic · May 28
Anthropic raises $65B at a $965B valuation
Anthropic raised $65 billion at a $965 billion post-money valuation — nearly tripling February's number and passing OpenAI ($852B) as the most valuable AI startup. Led by Altimeter, Dragoneer, Greenoaks, Sequoia. Run-rate revenue crossed $47 billion this month. Announced the same day as Opus 4.8.
The raise and the model dropped the same day on purpose — capability and capital are one story now. The number that matters for builders isn't $965B, it's the $47B run-rate: this is no longer subsidized research, it's a business with the margins to keep cutting prices on you.
"At a $965B valuation on $47B run-rate — is Anthropic priced like a software company or like a new utility, and does that change what you'll build on top of it?"
Cognition · May 27
Cognition: 90%+ of our own code is now written by Devin
Cognition (maker of Devin) raised $1B+ at $26B. CEO Scott Wu: more than 90% of the company's internal code is now committed by Devin — up from ~13% in December 2025. Revenue run-rate grew ~13× in a year to ~$492M.
The most important number of the week, and barely anyone's discussing it. A company that builds a coding agent now writes 90% of its product with that agent — 13% to 90% in five months. Either the proof-of-concept for the whole thesis, or the most spectacular dogfooding-as-marketing of the year. Either way, "what % of your code is agent-written" is now a metric people quote.
"What's your honest agent-written-code percentage right now — and what's the number where you'd stop being able to review it meaningfully?"
Anthropic · May 25
The Pope wrote an encyclical on AI — and invited the interpretability researcher
Anthropic co-founder Chris Olah was invited to speak at the presentation of Pope Leo XIV's encyclical "Magnifica humanitas." Olah's life's work is interpretability — literally "what is happening inside the model."
The technical frontier and the meaning-of-it-all frontier collided in one room. When the Church writes an encyclical on AI and seats the person who studies the model's insides, the question of who frames this technology morally stops being abstract. A good line to read aloud before we open the floor.
"When institutions like the Church start writing the moral frame for AI — who should be setting it: the labs, the institutions, or neither?"
Also in the bookmarks
Quick hits
Discussion · Demos · Q&A
What caught your eye?
The machine will happily do the work. — So what's the part you refuse to hand off?
Follow the lab.
See you next Saturday.