I have been testing personal AI agents across Instinct, Grok bot, Muse, and OpenClaw. Here are 5 observations from using them daily: 1. Context compounding is key An agent becomes significantly more useful when it holds deep context about your goals, habits, and preferences.

5. Adoption and consumer friction We are very early. Tinkerers use separate email accounts and sandboxes. Most people remain cautious. My spouse, for example, is cautious about granting email access. Solving that trust layer is what will make this mainstream.

DoorDash released a CLI and I couldn't not try it. I pointed Codex (GPT-5.6 Sol, High Reasoning) at years of my health data: blood reports, DEXA scans, WHOOP exports, plus my full DoorDash order history. Here's the Good, the Bad, and the Ugly:

The optimization function is still too narrow. The agent optimized for the stated goal and ignored goals it could have inferred from my own behavior. The data was already there.

Installed openclaw on sandboxed personal machine and excited to kick the tires. Already feel like a Luddite since I missed the first wave but as always, excited to try new tech and unlock its capabilities. At work, I am also very excited about other AI platforms catching up to Openclaw (e.g. Claude remote control and cursor automations). What a time to build! 💪

Very impressed w/ the AI tools at our disposal. All 3: Claude Code, Codex and Google's Antigravity are excellent! Also TIL: on my personal Google One Subscription, I have access to Claude Opus Model. Albeit it's one version behind.