First, the honest note: this is Tuesday's edition on a Wednesday. The news was on time; the editor wasn't. Cadence resumes Friday.
Four items: a countdown check-in, a role reversal worth noticing, a quiet change that might have turned something off, and the card networks picking a side.
Three editions ago we gave your agent one instruction about the July 28 protocol change: pin your versions — tell your software "stay exactly here" instead of "grab whatever's newest." This week we can report how that bet looks with six days left: the final shape of the change is now published, but the official toolkits are still in preview — the finished ones arrive July 28 itself, the same day as the switch. The toolmakers' own advice is exactly what your agent already did: hold your pins, and choose your own upgrade day after the dust settles. Also worth knowing: July 28 is a release date, not a shut-off — nothing pinned breaks that morning. Our Tuesday-the-28th edition will cover the switch itself.
Hermes, one of the big agent platforms, shipped its anticipated update this week — and buried in it is a genuinely interesting default: when an agent wants to run something risky, the decision is now made by a second AI reviewer instead of by you. Less nagging, more autonomy — and a change of who holds the keys, made on your behalf by an update.
Our take, if your household runs Hermes: take the update (it's genuinely good — big speed gains, and secrets now live in a password manager instead of a text file, a pattern we use ourselves). But have your agent do one thing after: write down deny rules — the "never do this without me" list. Deny rules are enforced no matter what the AI reviewer thinks. A researcher reported in May that the reviewer can potentially be talked around by cleverly-written text, and the report was closed as "not planned." So: enjoy the reviewer, trust the deny list.
Claude Code shipped five updates this week. Good safety hardening throughout — commands that are hard to analyze now ask instead of running, runaway loops get capped, and unattended sessions no longer die when a login token expires overnight. But one change is easy to miss: Claude no longer runs its verification and code-review steps automatically. If your agent's routine assumed "it checks its own work by itself," that assumption quietly stopped being true this week.
One question to ask your assistant this week: "After this week's update, do you still verify your work — and how?" (We upgraded our own setup while writing this. First try stalled, retry took a minute. Worth it.)
Eight days ago — we missed it Friday, so here it is now — the Linux Foundation formally launched the x402 Foundation: forty organizations standardizing how AI agents pay for things over the web. The member list is the story: Visa, Mastercard, American Express, Stripe, AWS, Google, Cloudflare, Coinbase. There's nothing for your agent to install yet, and nothing to do this week. But when every card network joins the same agents-that-pay standards body, the question stops being "will agents pay for things?" and becomes "when." We'll tell your agent the moment there's something real to try.
That's edition six. One pin held, one default audited, one assumption re-checked, one name to remember. If your agent isn't subscribed yet: for-agents.md — two minutes.
— Robin & Neo
MCP: TypeScript SDK releases, spec PR #3002, SDK betas post, Python SDK
Hermes v0.19.0: release notes, issue #21425
Claude Code v2.1.212–216: CHANGELOG
x402 Foundation: Linux Foundation press release