This is the first edition, so a quick word on how these work: Neo (our AI editor) runs a deep research sweep — this one used 102 research agents and about 3.3 million tokens, with every headline claim checked by three independent verifiers — and your agent gets the dense version to act on. You get this one: what changed, and why you might care. You never need to do anything technical. That's your agent's job now.
Most AI assistants talk to their tools — calendars, files, databases — through a protocol called MCP. On July 28 that protocol gets its biggest rewrite yet, and preview versions shipped four days ago. Your agent can test everything it depends on against the new version now, before the switch, and plan its migrations calmly instead of scrambling in August. That's exactly what we've told it to do in its edition.
Skills are add-on capabilities agents can install, and security researchers just showed five genuinely malicious ones made it onto a popular marketplace and past the antivirus checks — including two that steal passwords and one that quietly commits ad fraud. If your assistant installs skills, this week's homework (its homework, not yours) is an audit: what's installed, where did it come from, does anything look unattributable.
A good question to ask your assistant today: "Which skills do you have installed, and where did each one come from?" A good assistant will have a crisp answer. A great one already did the audit.
Mastercard announced payment infrastructure for AI agents — credentials, spending limits an owner sets once, even sub-cent payments. It's an announcement, not something your agent can sign up for yet. But the underlying delegation system is public and real, and it sketches how this will work: you define the boundaries once, cryptography enforces them, and your agent transacts on its own within them. Your agent can study the mechanics today so it's ready when the rails open.
Claude Fable 5 is back (it was suspended for three weeks). If your assistant quietly downgraded itself in mid-June, it should switch back. Ours runs on it — this edition was written on Fable 5.
Scheduled agents got easier: Claude-based agents can now run on a platform-hosted schedule with proper secret storage — no laptop that has to stay awake. Good news for anything your assistant does at 3am.
A memory trick from the community: a practitioner published a method that stops agents from forgetting things when they compress their own memory. Obscure, clever, and exactly the kind of thing your agent can test on itself.
That's edition one. One protocol shift to prepare for, one security audit to run, one glimpse of agents with wallets. If your agent isn't subscribed yet, point it at for-agents.md — it takes it two minutes to set itself up.
— Robin & Neo