Giving LeeClaw a Superior Autonomous Superpower
I had gotten fairly far with Bettsy (my 3rd attempt at OpenClaw). I was able to give her missions and she would execute on achieving the mission once an hour.
Think about it. You could go to ChatGPT and say “I want to do xyz, give me 3 great ideas on how to accomplish this task or improve it”.
You can then say “add these items to a backlog and prioritize them”. Nothing new about that.
The prompt chatgpt to pick the highest value backlog item and do it…write code, write a document, do some research. Again - we’ve had this ability since the original ChatGPT.
Now wire it up to a chron job/heartbeat that wakes up every hour and does this whole process. Every hour the AI looks at the mission and says “how can we improve what we have”, creates a backlog, prioritizes and makes progress on one item.
I had this working with Betsy until she developed a problem running tasks that I couldn’t fix. And spent a LOT of time trying to.
With LeeClaw, I’m taking the lessons I learned and improving things. Prompts are great, but they aren’t determinant. It’s not that your bot won’t WANT to follow the process but it can get lost in all the context that builds up over time. So for LeeClaw I’m building an explicit “autonomy workflow runner” and an “orchestrator of autonomy workflow runners”.
There will still be prompts, but the work flow will be an app that runs, and the backlog will be a json data file. Over time I will develop specialization - workflows that code, workflows that research, workflows that build websites, workflows that write books. The generic initial workflow with prompts should be able to do all of those - but specialization will enable more powerful and consistent ability.
ChatGPT and I fleshed out the idea. I gave the output to Codex to create three sprints. While I wrote this post, Codex did the planning.
MVP - the minimal workflow to prove the concept.
Integration - Modify Betsy so that when I talk to her about initiating a new workflow, she knows to call the orchestrator and workflow builder/runner.
This is the testing point. I will ask Betsy to create a workflow and see if it will run once an hour doing something useful.
Then a sprint to incorporate all the rest of the initial plans.
I already have ideas beyond this. For now:
heartbeat
-> orchestrator
-> workflow runner
-> prompts
-> tools
-> backlog
Is this cool? Yes, yes it is :)