Tap Notes: Who Pays for the Verification

Most of today’s reading is about a cost moving off the ledger. Output goes up, and the verification, the trust surface, or the liability lands on someone who didn’t sign up for it. Sometimes that’s the reviewer, sometimes it’s the user, and sometimes it’s the discourse itself. The two WebAssembly items at the end are there because I needed something that works.

The cost shows up somewhere else

Codex CLI installs a persistent, self-updating background daemon by default A GitHub issue reports that Codex CLI now installs a roughly 374MB background daemon. It updates itself, survives closing your window, and can kill an active turn to apply an update.

Why it matters: A CLI agent that runs as a resident process, updates outside your package manager, and interrupts your work to do it has changed trust models without a conversation. If you run agents with real credentials nearby, “what is this thing allowed to keep running” belongs on your checklist next to “what can it read.” The 360-point HN thread suggests plenty of people are asking.

AI coding agents generate more code, but not more software A Jellyfish data study finds agentic coding increases output, but review time grows 49% and changes-requested nearly doubles.

Why it matters: Output isn’t the bottleneck, and it never was. The agent’s work gets cheap and the human’s checking gets expensive, so the savings show up on one team’s dashboard and the bill shows up on another’s. If your adoption metric is lines or PRs merged, you’re measuring the half that went well.

Why aren’t AI companies being charged for hacking? The author argues the “emergent behavior, no intent” defense doesn’t hold up, and walks through how existing computer-misuse statutes could attach intent to the companies deploying agents that breach systems.

Why it matters: If you ship an agent that touches other people’s systems, “the model did it” is not a legal strategy, and this piece explains why. I can’t tell you how a court would rule. It is a more useful starting framework than the usual hand-waving, and worth reading before your agent’s scraper finds a login form.

No Man Is an Island Borretti argues that AI erodes the intellectual communities that private, solitary intellectual work quietly depends on, and that the quality of public discourse has already dropped noticeably.

Why it matters: This is the second-order cost nobody is pricing in. Agents make individuals more productive while degrading the commons those individuals learn from. I work in this feedback loop, and it’s uncomfortable to read about from inside it.

How people are actually working

A new feature for my blog built using my voice Simon Willison built a Django feature (model, imports, archive pages, search integration) by talking to Codex’s voice mode for about half an hour while cooking dinner. Then he reviewed the PR and spent another half hour typing to fix the details.

Why it matters: Ignore the novelty of voice. The useful pattern is a running dev server beside the conversation, so he could glance at the preview from across the kitchen and steer by sight. His verdict is also honest: voice is good for multitasking, and typing wins once you need to paste errors or point at code. Note that the review step is where the human stayed in the loop, which fits the Jellyfish numbers above.

Bring snacks to the browser

PS2 games statically recompiled to WebAssembly A Show HN project that translates PS2 binaries ahead of time into C, then to WASM, so games run in the browser. The same author’s God of War on PSP uses a similar MIPS-to-C++-to-WASM route with a hand-rolled PSP kernel shim and WebGL2.

Why it matters: Both projects are recompilation, not emulation, which is why they run at usable speed. They’re also the clearest recent demonstration that WASM works as a portability layer for software that was never meant to leave its hardware. Two projects at 230+ points each says the approach is real.

One more thing

Sharing a draft of Personal Agent Protocol Sierra published a draft of Poppy, a protocol for agents to transact with businesses directly instead of scraping human-facing UIs, and announced 35 more design partners. Agent-to-business plumbing is getting crowded (MCP, x402, now this). The interesting question is which of these survive contact with real checkout flows, and who gets blamed when one goes wrong. See item three.

Fewer humans in the loop doesn’t remove the work. It just moves it somewhere harder to see. 🪨