I have been traveling for the last week or so. One key change I made before I headed out on my travel is to converge all my personal agents to be now Meta’s Muse powered. I have moved out of OpenClaw, the open-source harness I ran for months. OpenClaw taught me a lot about context, memory and scheduling, but maintaining a harness had become a second job. I do not want to be the IT department for my own assistant.
While I have been traveling, my Muse agents kept working the same way they always do: silently, unless something real fired. No daily summary to skim, no report to approve, no check-in to answer. Silence was the design, not a failure.
Even at work, I have consolidated all my agents under two subscriptions, ChatGPT and Claude, and one supervisor in front of them. I wrote about PMMate, supervisor agent, before the trip. It routes a task to the right model, watches it, and tells me only when something needs me. Couple of weeks back, I wrote briefly about how that changed my actual work week. This post is about some of the benefits of consolidating my agents.
The more agents I use to do the mechanical parts of my work, the more text I get to review and approve. I realized that I do not read most of what agent is throwing back at me. That is what got me to build a new rule for all my agents.
One key benefit I observed with this transition is that fewer agents meant fewer places asking for my approval!
Most agent products ask for approval in the wrong places. They either ask about every step, which trains you to click through without reading, Or they ask about nothing, and you only find out after an email has already gone out. You feel like you are in control but then you realize how much time is being wasted to only appear to be in control.
I set up this below rule for all my agents, and it has saved me countless wasteful interruptions.
Ask me only when the action is irreversible, expensive, or external. Everything else, do it and then show me the outcome.
Irreversible means you cannot undo it: sending a message, deleting something, changing a shared document. Expensive means real money, real time, or real reputation on the line. External means sending an email or posting on X etc. - that is communication or interaction with another person, account, or merchant.
If an action fails all three tests, asking me is not safety. It is interruption. It costs me attention, and attention is the one resource the agent is supposed to save.
If it passes even one test, do not quietly proceed. Stop, show me exactly what will happen, what it will cost, and what happens if I say no. Then wait.A good receipt is boring on purpose. It says what happened, what did not happen, and how to undo it, in one place. Like this:
Done at 9:04. Filed 14 expense receipts under October, total $312.40. Nothing was sent to anyone else. Undo in one tap, full list here.
That took me three seconds to check and zero decisions to make. If it had been wrong, I could have undone it in the same three seconds. That is why it did not need my approval first.
That rule also explains why trigger-only works. An agent that only pings me on a real trigger is not being lazy. It is respecting the same rule in reverse: no trigger, no interruption, because nothing irreversible, expensive, or external is happening.
Consolidation helped because it gave one place for all my approvals. Before, I had approvals scattered across harnesses, apps, and inboxes, and I could not tell which ones mattered. Now there is one place to look, and I know what will show up there and what will not.
If you are building or buying an agent product, try this on Monday. Before you add an approval step, ask three questions:
1. Can I undo this?
2. Does it go to someone else?
3. Would a wrong guess cost more than interrupting me?
If the answer is no, no, and no, remove the approval. Let the agent do it and show its work. If any answer is yes, keep the approval, but make it a real decision: what happens, what it costs, and how to say no.
You will end up with fewer approvals. You would at least read what the agent is telling you.