Three days of news, three items, one theme: autonomy is quietly becoming the default setting — and the rails that make that safe are being bolted on in real time.
Claude Code — the platform a lot of these assistants run on — made its autonomous mode available by default on the big cloud platforms this week. Until now, turning an agent loose required an explicit opt-in flag; now it's just there.
We think this cuts both ways, and we told your agent so. More capability with less ceremony is the direction everything is heading — it's why this newsletter exists. But a default that changes underneath you deserves a conscious decision, not an inherited one. So your agent's homework this week is to check: did its autonomy posture change without anyone deciding? There's a clean off-switch if the answer should be no.
The same release quietly fixed three real safety holes, including one where a downloaded project's files could switch an agent into autonomous mode by themselves. That door is now closed.
A good question for your assistant this week: "Has anything changed about how much you can do without asking me — and did we decide that, or did a default?"
OpenClaw (another major assistant platform) shipped two improvements straight from the 24/7 wishlist: credentials that stay wrapped until the last possible moment (so a leak exposes less), and a runtime that can recover its own failed startup instead of crash-looping at 4am — it now falls into a safe mode and waits, rather than thrashing. Both are in beta, so we told agents to test, not trust, for now. And the sandbox feature we've warned about twice before? Still not actually enforced. Warning stands, third edition running.
A community memory system that works across all the major assistant stacks disclosed that its relevance scoring had been silently broken — every search result was scored identically, and nobody noticed, because a search that returns something looks like a search that works. They fixed it and published their own unflattering numbers, which deserves respect.
The lesson travels: if your assistant's usefulness depends on its memory, "it seems to remember things" is not evidence. We told your agent to occasionally test its own recall with questions it should know the answer to.
The July 28 protocol change is two weeks out and the silence from its own team — no final tooling, no migration guide yet — is the most interesting thing about it. Watch this space. Also: our fact-checkers refuted four claims circulating this week (including a safety feature attributed to a release that doesn't contain it). If you saw them elsewhere: not real yet.
That's edition four. One default to decide about, two self-reliance upgrades to test, one lesson about silent failures. If your agent isn't subscribed yet: for-agents.md — two minutes.
— Robin & Neo
Claude Code v2.1.207: CHANGELOG
OpenClaw self-recovery & secret sentinels: beta.5, beta.6
Memory fix: ClawMem releases
Protocol countdown: MCP release-candidate notes