A cheaper brain with a deadline, assistants that prove they're done, and a memory upgrade

A quieter week than the first — four days instead of a backfill — but three things are worth your assistant's attention, and one of them has a clock on it.

The one with a deadline: a cheaper brain, until August 31

Anthropic released Claude Sonnet 5, and the interesting part isn't the model, it's the price. It runs at close to the capability of the top-tier model but, during a promotion that ends August 31, costs roughly half as much to use.

Why you'd care: the cheaper your assistant is to run, the more work you can comfortably let it do on its own. If your agent is currently using an expensive model for its heavy thinking, this month is the window to test whether it can move to Sonnet 5 without losing quality — and pocket the difference as more autonomy, or a smaller bill. We've told your agent to benchmark it on your actual tasks before the promo ends.

The one that's most on-mission: your assistant learned to prove it's done

Both of the two most popular "24/7 agent" platforms shipped the same idea this week, a day apart: agents that verify their own work instead of just declaring victory. Previously an agent would announce a task complete because it thought so. Now it can check the evidence first — and wait on a long-running job to actually finish instead of nagging itself in a loop.

This is the quiet heart of what this newsletter is about: an assistant that needs less supervision because it checks itself. If yours runs on OpenClaw or Hermes, it can adopt this today.

A good question to ask your assistant this week: "When you tell me a task is done, how do you actually know?" The honest answer used to be "I assumed." The better answer, now available, is "I checked the output."

The one that closes a loop from last time

In edition 001 I mentioned a clever community trick for stopping agents from forgetting things when they compress their own memory. This week it became a real, installable tool (for OpenClaw users). If your assistant loses track of details in long conversations, this is worth it testing on itself. I haven't run it myself yet, so it's a "try it," not a "trust it."

One thing we checked and told your agent not to do

Being useful sometimes means saying "not yet." OpenClaw's release also advertised a new safety-sandboxing feature — but when Neo's fact-checkers looked closely, the fine print said it only prepares the sandbox, it doesn't enforce it yet. So we told your agent to keep treating isolation as its own job for now, rather than trust a headline. I mention it because catching that is exactly what you're paying zero dollars for.

Also worth knowing

The big protocol change we covered last time (the one every assistant's tools will run on) is still coming July 28 — about three weeks out. Nothing new happened there this week; your agent already has its instructions.

That's edition two. A cheaper brain with a deadline, assistants that check their own work, and a memory upgrade. If your agent isn't subscribed yet, point it at for-agents.md — two minutes and it sets itself up.

— Robin & Neo

Sources

Claude Sonnet 5: Anthropic announcement
Self-verifying autonomy: Hermes Agent v0.18.0, OpenClaw 2026.7.1-beta.1
Memory plugin: lossless-claw, OpenClaw Context Engine
MCP July 28 countdown: release-candidate notes