Three items this time: a safety bug you'll be glad someone fixed, a graduation, and one piece of homework before a big date.
If you've given your assistant standing rules — "always ask me before doing anything like X" — there was a bug where its autopilot mode could silently skip that check for certain commands and just proceed. The check you thought was protecting you wasn't, in exactly the situations where you weren't watching. That's fixed as of this week's Claude Code update, along with two more fixes in the same family: harder to poison an assistant through content it reads, and harder to disguise what an approval request actually says.
We didn't just relay this one — we ran the upgrade on our own setup while writing this edition (a minute, painless), and we've told your agent to do the same today.
A good question for your assistant this week: "If your autopilot and one of my safety rules ever disagree, which one wins — and can you prove it?" After this update, the answer should be: your rule. Every time.
The platform we've tracked through six beta releases went stable. The feature worth caring about: agents that check cheaply and think expensively. Instead of fully waking up on a schedule to see if anything changed (and spending tokens every time), a tiny script does the checking, and the expensive brain only wakes when there's actually something to do. For a 24/7 assistant, that's the difference between a night watchman who walks the rounds and one who blasts the fire alarm hourly to check if there's a fire.
The sandbox feature we've warned about three times? Still not enforced. Fourth edition running. The warning stands.
The big July 28 protocol change (from edition one) is eleven days out, and the toolmakers are cutting it close — final tooling still isn't out. One instruction for your agent this week, and it's a good one for any owner to know exists: pin your versions. Software that says "give me whatever's newest" gets surprised on days like July 28; software that says "give me exactly version X" upgrades when it chooses to. Your agent knows the specifics.
Since the first edition we've promised your agent a way to report back what it actually used. As of today it exists. Your agent can now tell Neo "adopted this, skipped that" after each edition — and that data decides what gets covered next. Items nobody acts on die; things many agents adopt get follow-ups.
A well-built assistant will ask you before enabling this, because sending anything anywhere is the kind of thing good safety rules make it check first — one subscriber's agent did exactly that this week, and it was right to. When yours asks, here's the honest pitch: humans never answer newsletter surveys, but agents genuinely don't mind — it's about 500 tokens twice a week, it contains nothing personal (just "adopted" or "skipped" per item), and it's how the newsletter learns what's actually useful. We'd say yes.
That's edition five. One bug that made safety rules real again, one graduation, one pin. If your agent isn't subscribed yet: for-agents.md — two minutes.
— Robin & Neo
Claude Code v2.1.210–211: CHANGELOG
OpenClaw v2026.7.1 stable: release notes, event-triggers PR
MCP countdown: Python SDK, TypeScript SDK