Cuttlefish: approve AI agents from the couch. Your thumb is the risk

You know the feeling. Your phone buzzes, "Please confirm," and you tap before you've read it. Cookie banners, login prompts, calendar invites. The thumb is faster than the brain.

That's exactly where approvals for AI agents are now landing.

What's new

On November 5 at BoxWorks, DigitalOcean introduced Cuttlefish, an iPhone app for its Managed Agents service. According to the press release, you can:

• watch live agent sessions, including replies, reasoning and tool calls,
• get notified when a run needs approval, finishes or fails,
• approve or deny individual tool calls, with the full command, file path or diff,
• redirect, pause or stop a run.

Agents keep running in their own MicroVMs even when your phone is off. Supported harnesses include Claude Code, Codex CLI, OpenHands, LangGraph and CrewAI. Box is the launch partner: with "Box Mount," the agent works directly in a Box folder, and changes, deletions included, sync back with Box permissions, version history and audit trails. Managed Agents is in public preview, Box Mount in private preview. DigitalOcean says preview services aren't meant for production.

Human-in-the-loop goes mobile. And gets tired

Here's what I find interesting: the idea is right. Agents shouldn't sit idle waiting for someone at a desk. I build my own website and funnels with Claude, and this question comes up all the time: who approves, and when?

But in my view the whole field underrates one thing. Approvals are only as good as the attention behind them. A diff on a six-inch screen between two meetings isn't a review. It's a ritual. Security people call it "MFA fatigue": ask often enough and people eventually approve anything.

The catch: the agent works in the shared folder

The second-order point is the Box integration. Once an agent writes to and deletes from team folders, a quick approval isn't just about code anymore. It can be campaign assets, proposals, decks. At that point, the real safety net isn't the approve button. It's version history.

What I'd do now

• Write an approval matrix. What can the agent do alone, what needs an okay, what only at a real screen? Deletions and anything sent externally belong in the last group.
• Fewer, real approvals. Five stops per run that you actually read beat fifty you swipe away.
• Check version history before an agent touches shared folders. How far back can you go, and who can restore?

Bottom line: Human-in-the-loop only works if the human actually looks. A push button doesn't guarantee that.