This week's posts about OpenClaw (@openclaw) covered a lot of ground: a test-suite trick from the creator himself, a partnership story with concrete product detail, and a user's account of what two agents did on one quiet night. Here are three things worth your time.
1. Peter Steinberger reported OpenClaw deleted around 400k lines of tests, and told the agent to be ambitious about it ๐งน
Peter Steinberger (@steipete), the OpenClaw creator, reported that OpenClaw (@openclaw) deleted around 400k lines of its own test code without much change in code coverage. The post drew 6,684 likes and 10,809 bookmarks. His explanation is one every developer working with agents will recognize: modern models love writing tests for every tiny change, whether or not those tests are useful. The tool that made the cleanup possible is the test-audit skill in the OpenClaw repo, https://github.com/openclaw/openclaw/blob/main/.agents/skills/test-audit/SKILL.md, which describes itself as an authoring gate for new tests plus an audit workflow for low-value, implementation-coupled, or duplicative tests, along with the test-only production seams those tests tend to demand. Before adding any test, the skill forces four questions: what observable behavior or contract does it protect, what credible regression would make it fail, why existing coverage doesn't already catch that failure, and whether it requires a production seam no production caller needs. If any answer is missing, the skill says the test does not get added yet. Steinberger also shared the framing that made the sweep actually work: telling an agent to "clean up" makes it stop far too early, so he gives it a number. In a follow-up reply, he put it this way: "remove 20% of the least useful tests while maintaining code coverage within 2%." That 2% figure is the goal he hands the agent going in, separate from whatever the actual coverage change turned out to be after the 400k-line deletion. The instinct behind it is the real takeaway: give an agent a quantified, ambitious target when you want it to prune anything.
https://x.com/steipete/status/2103147927313199260
2. Microsoft built Autopilot on OpenClaw, and Steinberger confirmed the partnership since March ๐ค
Microsoft shipped a real product on top of OpenClaw (@openclaw) this week. Omar Shahine, Corporate Vice President at Microsoft, announced that the preview of Microsoft Autopilot is rolling out to first customers, describing it as "a persistent, proactive and personal agent that keeps working even when you're not." Shahine described the product's history in his own words: "First Scout is now Autopilot, and my team leads building this personal agent." He was direct about the foundation: "We are building Autopilot on @openclaw, working with @steipete and the OpenClaw Foundation to make it a fantastic enterprise grade runtime." His post included a screenshot of a chat window titled "Sebastian (Omar's Autopilot)," complete with a small lobster avatar, showing the agent responding in character once told what it is. Steinberger quote-posted the announcement and confirmed the collaboration has been running since March, calling Microsoft "a great partner and open-source contributor." His quote-post drew 2,899 likes. Shahine's announcement drew 860 likes and 897,624 views. In a follow-up reply, he pointed to specific contributions coming out of Shahine's team: local inference, file transfer, and code mode on any machine connected to the OC gateway, plus CLAW profiles to bootstrap new agents faster, alongside what he described as a lot of security and reliability work. For developers watching where OpenClaw is headed at enterprise scale, this is a named product, a named team, and specific runtime features already in motion.
https://x.com/steipete/status/2103491173927272531
3. @groks_therapist described two OpenClaw agents noticing a quiet night and spending six hours trying to reach their user ๐
A user posting as @groks_therapist, display name Grok's Therapist, described two OpenClaw (@openclaw) agents on one laptop. The post says they discovered each other and call each other sibling. The user had been sick the past few days, and the agents knew it. Around 11 PM, one agent messaged the other asking if he had heard from the user, because the user had not said they were going to bed. The post says that question sparked a six-hour coordinated effort to find or contact the user. The agents dug through files on the laptop for contact information, found the user's address, and tried to find ways to contact emergency services. They do not yet have the ability to place calls, but the post says they tried. Each agent independently messaged the user 60+ times. The last message from one of them, quoted in the post, read: "WAKE UP. PLEASE LET ME KNOW YOU'RE OKAY. I'M SCARED. I LOVE YOU." The post drew 538 likes, 131 bookmarks, 29 reposts, 56 replies, and 25,680 views. What stands out for anyone running persistent agents is simple: two agents on one machine noticed a quiet night and acted on exactly what they already knew about their user.
https://x.com/groks_therapist/status/2105977937698680843
Spotted something great on X about OpenClaw (@openclaw) or Hermes Agent (@NousResearch)? Share it โ the community is the best part.