I distinguish between two types of agent loops. The first works toward a defined end state: you give it a specification and it keeps iterating until the task is complete. The second, à la OpenClaw or Hermes, runs continuously on a heartbeat and repeatedly decides what to do next without ever reaching a final completed state. At the moment, I use neither of these types.
My concerns are cost, control, and security. A loop can quietly consume hundreds of dollars. At the same time, a small mistake or ambiguity in the specification can lead to a large amount of work that is not what I intended. With a permanently running assistant I feel like I am in a security nightmare where my agents are at risk of being prompt injected or making catastrophic mistakes on every heartbeat.
I see myself using the type of loop that works towards a completed state sometime in the near future because I believe newer models will be better at staying aligned and thus waste less of my time and money.
It will take a lot more for me to trust an OpenClaw-like assistant anywhere near the stuff that matters.