I've become pretty agent brained over the last 72 hours. Here are some observations. What makes @openclaw different: 1. Self-modification, the fact that it can edit its own prompts, let's you collaborate with it to enhance it and configure it. In the field learning has eluded most tools to date. (also a huge source of security exposure) 2. HEARTBEAT and CRON tools. The fact that it has easy access to both schedule things, and just a general awareness loop. It's what makes it an agent and not just a parrot. 3. The ability to build more agents, the subagent paradigm is really powerful, maybe it's not something that survives the next 10x'ing in context window, but for now allows for some really powerful things, and creates good security boundaries. What it needs yet: 1. More context persistence, there are lots of tricky ways that you lose context. Calling into a subagent, getting a callback from a webhook, etc. You maintain agent level coherence, but any in context from the session that initiated the action is lost. There are work arounds but you kind of have to understand the moltbot architecture to use it correctly 2. It's missing a spark. The heartbeat is good, and there is some temperature in there, but I think it needs a spice mod, that just causes it to do random things once in a while. Humans do this all the time and I think it's a huge aspect of our general intelligence. I think this can be simulated, gonna try some ideas.