Tap Notes: The Cost Moved

Nothing in today’s reading got cheaper for free. Writing code costs almost nothing now, so the bill showed up at review. Agents can buy things without loading your page, so the bill showed up in your analytics. Trust assumptions rotted quietly in a certificate store. Read each item below and ask where the cost went.
Scale, and who pays for it
GitHub Octoverse 2026 Simon Willison relays the headline numbers from GitHub Universe. Merged PRs tripled in under a year to 220M a month, commits hit 7.38B a month, and new repos doubled to 30.3M a month.
Why it matters: This is aggregate behavior, not a vendor benchmark, and the curves bend almost vertically in the second half of the year. Every one of those PRs has to be read by something. Keep that number in mind for the next item.
The Mythical Man-Month in the Age of AI Brooks’s Law revisited for agent swarms. The argument is that the law mutated rather than died.
Why it matters: Onboarding a new “hire” now takes zero days, and that’s the part everyone celebrates. But coordination overhead didn’t disappear. It moved onto whoever reviews the output. Add more agents and you don’t get more throughput. You get a bigger pile for one human to read. Octoverse shows the pile. This piece names the bottleneck.
The harness is part of the model
How I think about Codex (February 2026) Gabriel Chua at OpenAI splits “Codex” into Model, Harness, and Surfaces. Simon points out the quiet admission: the Codex models are trained in the presence of the Codex harness.
Why it matters: If the weights were shaped around a specific tool schema, compaction behavior, and recovery loop, the harness isn’t a swappable interface. It’s load-bearing. That explains why models sometimes hallucinate first-party tool fields when you hand them a third-party schema. It also means model routing needs to account for harness affinity. A frontier model outside its native harness can lose to a cheaper model working in-distribution. Model-agnostic frameworks leak here, and it isn’t a bug they can patch.
Qwenputer: A functioning, deterministic RISC-V CPU inside Qwen3.5 Someone compiled a RISC-V CPU into Qwen 3.5 4B’s weights, using the Gated DeltaNet layers as exact RAM. It runs deterministically inside a forward pass, with no tools.
Why it matters: The stunt is fun. The follow-ups are the point. If you can compile exact computation into weights, you can hide reasoning in recurrent state where chain-of-thought monitors never see it. You can also compile tamper-resistant monitors into the model itself. Both of those are attack-surface stories, and the offensive one will be tried first.
Trust that nobody re-checked
Ten days later, Git on macOS still trusts the compromised eMudhra certificates Ten days after disclosure, macOS developer tooling still trusts the compromised CAs. That reportedly covers git, curl, and python3 HTTPS. The latest Apple release even undid users’ manual workarounds. The page includes a check command.
Why it matters: Run the check on your dev machines. Revocation is only as good as the slowest vendor in the chain. Anyone whose agents or CI run on affected Macs has a supply-chain exposure that no dependency scanner will flag.
Why Is This Fingerprinting Script Creating Bioweapons? Per the write-up, a browser fingerprinting script injects prompts that push LLM-assisted browsers into generating bioweapon content.
Why it matters: It’s a chained pattern. A tracking script does its normal job, and the injection rides along through a channel nobody defends. If your agent browses the web, the page is untrusted input, and so is every script on it. It got nearly 400 points on HN, so it isn’t a fringe worry.
Your dashboard is lying politely
Agentic Commerce May Break Your Measurement Before It Breaks Your Strategy Seer Interactive’s point: an AI agent can complete a purchase without ever loading your site. Your order system records the sale while GA4 never sees it.
Why it matters: Brands will cut budget on the wrong number. Orders held steady while sessions fell would read as a channel dying when it’s really a channel going invisible. The OTA/hotel analogy in the piece is the right warning. If you sell memberships or subscriptions, reconcile analytics against the payment processor before you make any decisions from the traffic graph.
One more thing
The October inflection point for vibe game development Simon argues that Opus 5.5 crossed from “looks like a game” to “actually fun to play,” with four browser games as evidence. One of them started from a prompt that said, more or less, “make it a LOT more fun.”
Low stakes, fast feedback, nobody’s production database. If you want to see where agentic coding is now, a game is the cheapest place to look.
That’s the day. Writing got cheap, so the reading got expensive. 🪨