visible infrastructure
self.md radar — 2026-04-16
today looked less like smarter chat and more like AI turning into visible infrastructure: browser-side artifacts you can inspect, public governance splits you can point at, and robot factories that suddenly have throughput numbers.
lead with the browser as both runtime and capture surface, pivot to governance getting concrete instead of abstract, and close on physical-world automation to widen the frame.
1. the browser stops vibing and starts leaving artifacts
sources:
what happened: libretto pitches browser automation as deterministic and inspectable rather than live prompting against a DOM. a Claude Code plugin now eats a live site and spits out a reusable markdown design system plus a Tailwind config. separately, a 1-bit Bonsai 1.7B at 290MB is running locally in-browser on WebGPU.
why this matters: the browser is quietly becoming both execution target and extraction surface, and the more useful agents are the ones leaving behind scripts, tokens, configs, and local runtimes you can actually open and read.
2. anthropic breaks ranks on the liability shield
sources:
what happened: Anthropic publicly opposed an Illinois bill, backed by OpenAI, that would shield labs from liability for large-scale harm. in parallel, users are starting to see identity verification enter the Claude access conversation — not a universal rollout, but enough that it reads as account policy rather than a blog-post ideal.
why this matters: governance is moving out of abstract alignment language into liability carve-outs plus access gates, and the real fight now is over who carries downside and who gets through the door.
3. humanoids get a factory cadence
sources:
what happened: Leju Robotics says it has an automated humanoid factory producing one robot every 30 minutes. the story isn’t the walking-demo footage — it’s the shift from prototype reveals to something that sounds like line speed.
why this matters: once humanoids have a throughput number, the metric stops being demo virality and starts being cadence, and software-style iteration loops begin leaking into hardware labor markets.
supporting links
- inkeep/agents — visual builder plus TS SDK, the cloud-agent flavor of today’s artifact shift.
- open-agents — Vercel’s open template pushing agents from chat toy to deployable surface.
- hermes-agent — personal-agent shell worth watching; evidence still too thin for main billing.
- decoder blocks training video — side-door into the same make-model-behavior-visible theme.
left on the table
- Markdown Kanban | No Plugin — already shipped yesterday, rerunning would be lazy.
- Anthropic Opus 4.7 rumor — rumor-grade, and too close to recent Anthropic-heavy coverage.
- anti-vibecoding tool for Claude Code — same bucket as yesterday’s anti-vibes signal, weaker example.
- OpenSandbox — resurfaced old repo, no fresh delta.
- Claude KYC thread — potentially important, but still too screenshot-rumor-shaped to anchor its own section.