Living on the Frontier · Session 9 · Online
Unattended
April 18, 2026 · The AI Lab, Chiang Mai
They shipped a model that verifies its own work.
They shipped a harness that runs without your laptop open.
They shipped a loop that paces itself.
The desk pet lit up amber.
The valuation swelled past eight hundred billion.
Meanwhile, the protocol everyone built on
shipped insecure-by-default, by design.
The agents are leaving your sight.
This week
Part I
The Handoff
A week of features that run without you watching
Anthropic · Apr 16
Claude Opus 4.7 — hand off your hardest work
Anthropic's most capable Opus yet. SWE-bench Verified 87.6% (from 80.8%), CursorBench 70% (from 58%), GPQA Diamond 94.2%. New: xhigh effort, task budgets, /ultrareview multi-agent code review, vision up to 2,576px (3.75 MP — 3x 4.6). Ships on Bedrock, Vertex, GitHub Copilot day one.
The phrase that matters is "verifies its own outputs before reporting back." For a year the dominant failure mode of coding agents has been "confidently wrong." Baking verification into the model's loop (not an external harness) changes the trust contract. If it ships as advertised, you can leave the room while Claude works — which is the whole bet of auto mode.
"At 87.6% SWE-bench, how much of your daily coding is now 'the model can do this unattended'? When it can't, what's the bottleneck — model, harness, or prompt?"
Anthropic · Apr 14–16
Auto mode + Routines — no prompts, no laptop
Two features, one direction. Auto mode (Max users) kills permission prompts — hand Claude a task, come back to verified work. Routines (research preview) bundles a prompt, a repo, and connectors into a single config that runs on a schedule, via API, or from a GitHub event — on Anthropic's cloud, not yours.
Auto mode is the tier-segmentation play: pay $200/mo, the agent runs unattended. Routines is the infrastructure play: your coding agent lives in the cloud, triggered by your git events. Combined, they're the same bet — the future of coding is tasks that run while you sleep, not interactive pair-programming.
"If your coding agent runs on someone else's infra 24/7, what's the right way to think about cost — per-token, per-routine, or 'a junior dev equivalent'?"
Anthropic · Apr 15
Claude Code desktop — one window, many sessions
Rebuilt desktop app with a sidebar for every active and recent session. Filter by status, project, or environment. Integrated terminal for tests/builds, in-app file editor, rebuilt diff viewer for large changesets, expanded preview pane handling HTML and PDFs alongside local app servers. Drag-and-drop layout.
This codifies what power users have been doing with tmux for a year — one session per task, running in parallel. The single-terminal "pair with one Claude" model is no longer the UX. Anthropic is telling you the right mental model for 2026 is "fleet of agents you dispatch and check on," not "assistant you chat with."
"If you're running 5 Claude sessions in parallel, what's the right ratio of your attention per session? Is 3 the actual limit before quality drops?"
Anthropic · Apr 10, Apr 16
Dynamic /loop + Bluetooth API — self-pacing + ambient signal
Dynamic /loop: run /loop without an interval and Claude picks the tick cadence based on the task — or skips polling entirely via the Monitor tool. Bluetooth API: Claude Cowork + Code now talk to hardware. Felix Rieseberg built a desk pet that lights up when Claude needs permission.
Tiny features, big direction. Self-pacing is the sleeper UX — too-fast loops waste tokens, too-slow miss signals; letting the model decide is small change, big cognitive relief. Bluetooth is the physical surface for "your agent needs you." Both only matter once agents run long enough that you're not watching — which is where the product is heading.
"What hardware signal would you actually want from your agent — blinking LED, buzzer, colored-light 'state', audio cue? Or is this a solution to a problem you don't have?"
OpenAI · Apr 15
OpenAI Agents SDK — open harness, BYO sandbox
Major SDK update: build Codex-style agents with primitives for file/computer use, skills, memory, compaction. Harness and compute split — BYO sandbox or use Blaxel, Cloudflare, Daytona, E2B, Modal, Runloop, Vercel. The harness is open source.
OpenAI's "we surrender on closed harnesses" moment. Anthropic wins developer mindshare with Claude Code precisely because the harness is inspectable. Now OpenAI open-sources theirs too. The real product isn't the SDK — it's the 7-partner sandbox list. If you're building agent infra, you're either on that list or outside the approved ecosystem.
"If the harness is commoditized (both Claude Code's and Agents SDK's are open source), what's the actual moat — model quality, sandbox partnerships, or developer habit?"
Part II
The Crack
Unattended also means unsupervised
Ox Security · Apr 15
MCP: systemic flaw, 150M downloads
Ox Security research on a "critical, systemic" flaw in MCP's STDIO execution model — baked into Anthropic's official SDKs across Python, TypeScript, Java, and Rust. Enables arbitrary command execution: API keys, chat histories, internal databases exposed. Anthropic confirmed the behavior is by design, declined to change it, and stated sanitization is the developer's responsibility.
A governance moment, not a bug moment. Anthropic shipped MCP, gifted it to the Linux Foundation, and now defers security boundaries to every downstream implementer. Defensible in protocol theory; brittle in practice — 150M downloads of vulnerable SDK code because defaults ship unsafe. "Sanitization is the developer's responsibility" is technically true and historically the origin story of every major supply chain attack.
"When a protocol is 'secure by design' but reference implementations ship insecure defaults, who owns the CVE — protocol authors, SDK maintainers, or every app developer who adopts it?"
Part III
The Stakes
Money, model gating, and who sets the rules
Bloomberg · Apr 14
Anthropic at $800B — Berkshire territory
Multiple investor offers at an $800B+ valuation for a new funding round. The Long-Term Benefit Trust also appointed Vas Narasimhan (outgoing Novartis CEO) to the board on Apr 14.
$800B sits between Berkshire Hathaway and Walmart — for a company whose primary revenue is still API token sales. The math only works if Claude captures 30%+ of the agent economy Anthropic is actively building the standards for (MCP, AGENTS.md, AAIF). Narasimhan's appointment is a tell: pharma execs understand regulated, high-stakes deployment — which is where the margins sit once coding tools commoditize.
"Is $800B a bet on Claude, on agents, or on Anthropic's ability to set the rules for the whole ecosystem? What's the Year-5 revenue story?"
Bloomberg · Apr 16
How Anthropic decided Mythos was too dangerous
Bloomberg's deep dive on Anthropic's internal decision not to release Mythos publicly. The Apr 7 Mythos Preview was the controlled version — enough capability exposure to launch Project Glasswing (defensive cybersecurity) without letting the full model out.
First public worked example of a frontier lab doing capability gating at the model release level, not the deployment level. The frame: "the model exists, but only we get to use it, and we'll use it for defense." Responsible stewardship or capability hoarding — depends on your priors. But it's a precedent; if Anthropic can sit on Mythos, other labs will point at this moment when they hold back their own.
"Is 'we built it but we're not releasing it' a responsible posture or a market one? Who audits that decision, and what would make you trust the audit?"
Also in the Bookmarks
Quick hits
Discussion · Demos · Q&A
What caught your eye?
"When was the last time you sat and watched an agent do the whole task?"
Follow the lab.
See you next Saturday.