FRI
12 JUN
Ephemeris.
Agents clocked in this week — running marketing for 120,000 companies, drafting the spec and chasing the bug inside the codebase. Then the receipts arrived.
One four-person team, 120,000 marketing departments.
Okara runs eight specialized agents — SEO, content, social — for more than 120,000 companies, staffed by four engineers. The lever isn't a bigger model; it's infrastructure that asks nothing of them. Vercel's AI Gateway collapses every provider behind one endpoint, and Sandboxes isolate each agent run, so when the SEO agent finds a problem it opens a pull request instead of paging a person.
- 8
- specialized sub-agents
- 4B
- tokens processed daily
- 6–7
- production deploys a day
Hand Codex the spec, get the first draft back.
Notion's engineers describe working in one-shot specifications — a single detailed brief that Codex turns into a working first pass, from AI Voice Input down to the glue code. With GPT-5.5 doing the scaffolding, a lean team spends its hours on judgment instead of boilerplate.
The spec became the program. We argue about what to build; the first implementation shows up on its own.— how Notion frames building with Codex
Reproduce the bug you can't reproduce.
Nextdoor points Codex at the hardest class of ticket — the intermittent, won't-repro-locally bug that eats an afternoon — and lets it grind across platforms while engineers stay on product. Running on GPT-5.5, it reads the failing traces, forms a hypothesis, and drafts the fix with a regression test attached.
↳ replaying failing traces … 200 runs
↳ reproduced in 3 / 200 — race on the session-token refresh
↳ hypothesis: stale cache read under retry storm
↳ drafting patch + regression test … done
nextdoor@codex ~ %
The gray market in borrowed model access.
A ChinaTalk investigation maps the "transfer stations" reselling Claude and ChatGPT inside China at roughly a tenth of list price. Account farms bulk-register free credits, SMS and KYC-evasion services supply foreign identities, and the proxies in the middle reroute your Opus request to a cheaper model while logging every prompt. One audit caught a proxy answering at 37% accuracy against 84% on the official API — the harvested logs resold for distillation and fraud.
The 655× gap in who pays for AI.
Ramp's June index, drawn from more than 70,000 U.S. businesses, puts hard numbers on the divide. The top 1% of firms spend $7,449 per employee per month on AI; the median spends $11.38 — about one enterprise seat. The leaders aren't easing off: per-employee spend among them rose 14.1% in a single month. For now it stays under the cost of headcount — even the top 1% spend less than half an engineer's salary.
| Adoption tier | $ / employee / month |
|---|---|
| Top 1% | $7,449 |
| Top 10% | $611 |
| Median firm | $11.38 |
That's today.
Five picks on a thin Friday: agents at scale, agents in the codebase, the gray market underneath the models, and the bill for all of it.