AI agents with command limits and approval checks

Results

The problem

I use agents to read logs, draft scripts and write up findings while I'm away. They can misjudge a command even with careful instructions, so I added command checks outside the model and an approval step.

What I did

• A chat interface that starts a headless Claude Code session for each message. A gate prompt sorts requests into do it, ask me first, or never.
• A deny hook checks shell commands against fixed rules for the router, production cluster, file server and disk wipes. A match blocks the command regardless of the model's decision.
• An email assistant whose read runs cannot send mail. Sending or archiving requires my approval of a specific command ID in chat.
• Claude and GPT review the same plan against a shared rubric and provide evidence for their scores.
• Written operating rules, backed by separate code checks where a mistake would be irreversible. I review and update the rules when a test exposes a gap.