Two months ago I asked here how people stay aware of what their agents are doing, and @getosmo replied with something that has been quietly rearranging my roadmap since.
I had been treating "can you undo it" as the line for when an agent should stop and ask. He took that apart. Irreversibility over-fires and under-fires, he said. It over-fires on things that cannot be undone but that nobody cares about, like the agent writing a log line. It under-fires on the ones that really hurt: a mass email to 50k users, or a migration you can only roll back after downtime. "Reversible on paper, ruinous in practice."
When an AI agent stops to ask permission, Pushary sends the question to the person running it, in their Mac’s notch or on their phone. Lee’s agent asks Lee. Sam’s agent asks Sam. A rule can send a push to your infra repo to your lead instead. Answers from the phone, Mac app and dashboard are saved with a name, so when something breaks, one page shows whose agent asked and who said yes. Works with Claude Code, Codex, Cursor, Agy and Hermes.
After enough "can I run the tests?" prompts, "can I send this to your customer?" risks becoming another button to clear.
Five months ago I asked here how people keep track of their agents. The replies went well beyond coding. In a later discussion, a builder running customer support agents made a point that stuck: people stop reviewing when checking the draft takes longer than writing the reply themselves.
Auto mode is the default in Claude Code now. A classifier checks every tool call, runs the safe ones and blocks the risky ones. Most days that is exactly what you want.
The part I keep coming back to is what happens after a block.
I had Claude Code running behind a notes window. It reached a permission prompt and sat there while I kept working. I was on the same Mac the whole time.
That felt worse than missing a finished task. I was not away. The terminal was simply not where my attention was.
one of the shops we support does photo prints. someone asked if they also print diplomas. they do not. the assistant gave a vague answer, the customer asked again more directly, and the second time it said yes. they sent the file to print.
we sorted it out and nobody lost money, but the second answer is the part i keep thinking about. the first one was hedged. it only became a yes after the person pushed. so it did not really invent a product, it agreed with a confident customer.
I run Claude Code, Cursor, and Codex most of the day, and I keep flip-flopping between two setups that both feel wrong.
Setup 1, babysit: I watch every command and approve everything myself. Nothing slips past me, but I'm stuck at the terminal and the whole point of running an agent is gone. I might as well be typing it myself.
One command connects your AI agents to your lock screen. When Claude Code, Codex, Cursor, Gemini CLI, Hermes, or Claude Cowork stops to ask, you tap yes from your phone and the run keeps moving. New in this launch: native iPhone and Android apps. Pairing is a QR code you scan from the terminal, no key to paste. Per-tool policies auto-approve the safe reads. Every decision lands in an audit trail.
