Living on the Frontier · Session 10 · In-person
The Arms Race
April 25, 2026 · The Kannas Hotel, Chiang Mai
Codex shipped memory that reads your screen.
Codex shipped review that runs while it works.
Codex shipped a browser and told itself to click.
Four million users in a fortnight.
A new model called “telepathy” from the inside.
Claude shipped a fleet to hunt bugs in the cloud.
Claude shipped live artifacts that refresh themselves.
Claude shipped the recap for when you context-switch away.
And somewhere underneath, Vercel broke open.
Nobody’s slowing down.
This week
Part I
The Race
Codex vs Claude Code, shipping weekly
OpenAI · Apr 23
GPT-5.5 — pitched at agents, not chat
"A new class of intelligence for real work and powering agents, built to understand complex goals, use tools, check its work, and carry more tasks through to completion." Shipping in ChatGPT and Codex, all paid plans. API coming soon.
OpenAI shipped a half-version bump pitched specifically at agents: "check its work, carry more tasks through to completion." They're no longer marketing to chat users, they're marketing to builders — the same positioning shift Anthropic made 18 months ago. Sam's internal codename was "telepathy."
"Does GPT-5.5 change your default model for agentic work? If you switch, what's the actual decision trigger — benchmarks, cost, or vibe?"
OpenAI · Apr 21
Codex hits 4M weekly users — +1M in 2 weeks
"Codex is now used by more than 4 million people every week." 1M users added in under two weeks. To celebrate, they reset rate limits again.
1M users in 14 days is hockey-stick territory. Pair it with JetBrains' data that Claude Code and Cursor are tied at 18% workplace adoption — Codex has distribution (default-enabled for ChatGPT Plus), Claude Code has craft. The question is which moat holds. Historically, distribution wins.
"Is Claude Code winning on craft and losing on distribution? And does that matter if the work still gets done?"
OpenAI · Apr 24 — today
Codex browser use — the build/verify loop closes
"You can ask Codex to build your front end, and test it like a user would by clicking through the app. Codex sees everything a user sees through vision & checks the network/console logs to help debug & fix any issues."
This might be the most important ship of the week. The build-verify loop was the last thing humans did that agents couldn't close on their own. Now Codex generates UI and clicks through to test it. If this works, the delta between "Codex writes code" and "Codex ships features" just collapsed. Watch the demos critically though — Playwright-in-a-trench-coat is not the same as real user testing.
"When the agent can see its own output through the eyes of a user, does QA survive as a role — or does it mutate into 'adversarial test agent orchestrator'?"
OpenAI · Apr 20 & Apr 23
Codex Chronicle + auto-review — memory and autonomy
Chronicle (Apr 20): Codex memories now read recent screen context. The agent knows what you were looking at before you prompted it. Auto-review (Apr 23): Codex works longer with fewer approvals — a separate agent checks higher-risk steps in context before they run.
Two features, one direction: expand the attention surface (what the agent sees without being told) and expand the autonomy surface (how long it can run without being stopped). The privacy argument is real — an OpenAI-owned agent is now watching your screen. The UX argument is also real — you don't restate context.
"Would you enable screen-context memory in Codex if it made your workflow 30% faster? Where's your personal line?"
Anthropic · Apr 22
Claude Code /ultrareview — fleet of bug-hunters in the cloud
"/ultrareview (research preview) runs a fleet of bug-hunting agents in the cloud. Findings land in the CLI or Desktop automatically. Run it before merging critical changes — auth, data migrations, etc." Pro and Max get 3 free reviews through 5/5.
Anthropic's answer to "LLM reviews still miss real bugs" — throw a fleet at it. Compare to Codex's auto-review: OpenAI runs a reviewer inline during execution, Anthropic runs a swarm after. Different bets on where the compute goes. Both are solving "how much autonomy can we grant before something dumb happens."
"Would you trust a swarm of cloud agents more than a single careful reviewer for an auth change? What would the swarm have to prove to earn that trust?"
Anthropic · Apr 20 & Apr 21
Cowork live artifacts + Claude Code session recaps
Cowork: Claude now builds live dashboards and trackers connected to your apps and files. Open one any time, it refreshes with current data. Claude Code: the terminal now shows recaps when you switch focus away from a session and come back. Built explicitly for multi-claude workflows.
Live artifacts is Anthropic's shot at Retool / internal tools as a category. If Claude spins up a wired-to-your-data dashboard in 10 seconds, Retool's $20/user/month pitch gets thin. And "multi-claude" being acknowledged as first-class UX is the story in the recaps update — six months ago that was a weird power-user hack. Now it's the spec.
"If you're running 5 parallel Claudes, what's the bottleneck — their output speed, your review bandwidth, or the context you lose switching between them?"
Part II
The Edges
Plumbing, platforms, and the shape of what's next
Anthropic · Apr 22
Anthropic blog: MCP vs APIs vs CLIs
"Building agents that reach production systems with MCP." When should agents use direct APIs vs CLIs vs MCP? Patterns for MCP servers, context-efficient clients, and — crucially — pairing MCP with skills.
MCP for tools that benefit from model-driven orchestration. Direct APIs when latency matters. CLIs when the tool is already one. The interesting move: Anthropic is framing skills as the higher-level abstraction and MCP as the transport. Which means if you're building agent infra, the architecture decision isn't "do I ship an MCP server" — it's "do I ship a skill, and does it need MCP underneath."
"If skills are the UX and MCP is the plumbing, what's the equivalent on the OpenAI side — and why are they so much quieter about their architecture?"
Google Cloud Next · Apr 22
Google Agentic Data Cloud — connective tissue for agents
Google announced the Agentic Data Cloud at Cloud Next: connective tissue so agents can traverse enterprise data without integration plumbing. The Data Agents Kit explicitly targets Claude Code and VS Code, letting developer environments orchestrate cloud resources autonomously.
Google is doing what AWS did with S3 in 2006 — betting data gravity beats protocol wars. If every agent reaches through Google's connective tissue, Google owns the audit log, the permissions, the meter. MCP is the protocol; this is the billing surface.
"Does MCP still matter if the big clouds turn agent access into just another IAM role?"
OpenAI · Apr 22
ChatGPT workspace agents
"Shared agents that can handle complex tasks and long-running workflows across tools and teams." The workspace part is what's new — agents are now team-level objects, not per-user.
OpenAI going straight at Notion AI / Slack AI / every "AI teammate" B2B pitch. Shared (not per-user) means the billing and governance model finally matches enterprise buying. This is the enterprise pivot showing up in product, not just sales.
"Does 'shared agent' solve the fundamental problem with AI-at-work, or just move the question of who gets blamed when it goes wrong?"
Part III
The Cracks
What breaks when everyone ships this fast
Breach · Apr 19
Vercel breached by ShinyHunters
ShinyHunters claims they've breached Vercel. Vercel confirms: "a security incident that involved unauthorized access to certain internal Vercel systems, impacting a limited subset of customers." The bulletin is out — specifics are thin.
ShinyHunters has a pattern: breach, post screenshots, real impact weeks later. The more interesting signal: every team in this room has code deployed on Vercel's infra. The blast radius of a breached deployment platform is different from a breached SaaS app — secrets, build artifacts, serverless function code, routing.
"When your deployment platform gets breached, what are you actually exposed to that you weren't thinking about — secrets, build artifacts, serverless function code?"
Anthropic · Apr 22
Mythos: the model too good to ship
Anthropic quietly released Claude Mythos Preview to a limited set of companies. 93.9% SWE-bench Verified. 94.6% GPQA Diamond. Found thousands of zero-day vulnerabilities across every major OS and browser. Paired with Project Glasswing to harden infrastructure before attackers catch up.
First time a frontier lab has withheld a model because it's too good at offense. The numbers matter less than the signal: we're in asymmetric-capability territory, where defenders need the same model class just to keep parity. If you build security tooling, the ground shifted this week.
"Is 'too dangerous to release publicly' the new moat, or a preview of a permanent two-tier world where the best models only go to governments and hyperscalers?"
OutSystems · Apr 22
Enterprise "agent sprawl" — 94% raise concern
New research: 54% of enterprises have integrated AI agents into core operations as of mid-2026. 94% raise concerns about "agent sprawl." 46% cite integration with existing systems as their primary deployment challenge. Microsoft reframes Copilot as Agent 365 with permission structures, audit trails, real-time monitoring.
"Agent sprawl" is the new "VM sprawl" from 2012. It predicts a wave of FinOps-for-agents tooling in the next 12 months. If you're building B2B, the pitch isn't "our agents are smarter," it's "our agents leave a clean audit trail." The governance-first pitch now scales faster than the capability-first pitch.
"How do you decide which problems deserve an agent versus a boring scheduled job? What's the test that keeps your own sprawl in check?"
Quick hits
Discussion · Demos · Q&A
What caught your eye?
If Codex and Claude Code are shipping every week, what happens to your own pace?
Follow the lab.
See you next Saturday.