The first month I used it like crazy. Now I use a third of it.
The first month felt like a new world. I connected every account I could and asked it to act whenever I could. As someone who builds agents, some of it inspired me, and some showed me the wall of the technology. There were things you cannot know without using it yourself. The feeling of an agent going into my accounts and doing my work for me cannot be explained. What inspired me was the connection: connect an account and the agent picks up my work. What showed me the wall came next. Connecting is easy. Finishing the connected work is another matter. First, the intelligence is low. At launch they said it was frontier-level. Now it does not compare to frontier. So I repeat commands, and re-command the parts I corrected. A job that should end once takes two or three rounds. Low intelligence makes commands longer, and longer commands mean more fixing. As an agent builder, this is the most disappointing part. Intelligence was what inspired me, and intelligence is what holds me back. With a frontier model, a short command is enough; it fills in the context itself. Now I fill in all the context myself, and it still sometimes gets it wrong. Then the command gets longer and the result gets shorter. Second, search accuracy is low. At first it was amazing that it searched through the browser like I do, and free on top of that. But the accuracy is low, and I think that is also about intelligence. It clicks the wrong thing, or brings back what it thinks is right. So search works, but the quality is low, about 30 to 40 percent by feel. The problem is not that search fails. It is that I have to re-check what it brought back. When checking costs more time than it saves, searching myself is faster. One search and two checks makes a search half a search. I reopen its links and re-check its numbers. At that point I might as well search from the start. So I let it search, but I search the important ones myself. Third, security may become a problem later. Meta's chief says they take security seriously, but my Facebook account was hacked recently, and the login log showed access records from many places besides me. Many of them were agent apps like Muse. But I cannot confirm the hack happened because of the agents. What I do think is this: for convenience, I am opening my accounts one by one, agents are multiplying, and whatever I open will be remembered and used somewhere, no matter how independent the space claims to be. Independent does not mean unremembered. I believe that if the convenience is there, security will improve. But right now the Muse app on Muse Spark interests me less, and the cycle of redoing work two or three times has grown. These days I keep searching for when Muse Wonder will arrive. I want to feel whether higher intelligence makes it better. Convenience brought me in. Intelligence will decide whether I stay. When your use of an agent dropped, what brought you back?