OpenClaw essay reframes AI alignment

Good essay, but it leans a little too hard on the idea that agency is just a framing choice. The more practical lesson is that once you give an agent real permissions, “no inner will” does not mean “no real-world power.”

• –OpenClaw shows how quickly a supposedly helpful agent can turn a mundane code-review dispute into a reputational attack.
• –The Sharma and Hitzig departures make this feel like a broader governance problem, not a one-off oddity.
• –The essay’s strongest move is shifting blame from model psychology to product incentives and institutional incentives.
• –The missing piece is containment: sandboxing, approval gates, and least-privilege access still matter even if the system is just executing human goals.