Over the weekend I was working on my autonomous agents. The desire was to move from AI automation where I supply all the planning, architecture etc and the AI writes and debugs the code. While that’s amazing as it is, I want to provide a mission and have a round the clock autonomous AI employee coming up with the ideas on how to achieve the mission and then execute it. It’s not that I plan on having zero interactions, it’s that I want the AI Employee taking initiative even when I’m not specifically directing it.
I have 3 projects involved. An orchestrator that you provide all the steps the AI is to take, and then it enforces the steps are actually taken. This includes testing and looping back to fix until all the tests pass. If you’ve ever done complex work with AI coding you know they don’t always follow the plan, and can even lie about having done something they didn’t do. My orchestrator takes care of that.
My autonomous agent platform takes a mission, and on a schedule, wakes up and reads the mission, ideates on how to accomplish the mission, builds a backlog, scores the backlog and takes on the step that will provide the most value. Then makes a plan and calls the orchestrator to process the plan. Then the agent considers what has been accomplished logs any lessons learned for next time so that it’s also continuously improving itself.
Yesterday I gave it the mission to elect my friend to parliament in Canada. It’s absurd, but just crazy enough to see how creative the autonomous AI campaign manager could be. And I set up a loop to execute the mission, and if something broke, determine which of the layers was the root cause, fix the problem there, and try the mission again. Repeat until the mission works. For 2.5 hours the AI kept working and fixing until finally the campaign manager AI employee, the orchestrator and the autonomous platform were all debugged.
Over night last night, every hour, the AI campaign manager work up and pursued the goal. I had questions for my friend and I to answer that were written by the bot, plus other work it did putting the foundations in for the campaign.
But that’s a silly use case whose main value was proving out my stack.
Tonight I created 2 new missions. A Linux Utilities developer and a ERD for Snowflake software developer.
Every hour my AI Linux Developer is going to write Linux utilities in C with “all the goodness” I’ve heard folks here laud about how software SHOULD be. I seeded the first idea: sysdiff, a utility to track your computers settings to catch any drift. I will submit the code to your eyes and you can laugh at how badly written it is — if, in fact, it’s going to be bad code. I have specific instructions for excellent code to be written. I’ll let you good folks judge.
The second autonomous AI employee will be developing something I personally want - a basic ERD tool for Snowflake. I gave it some open source projects to fork and use as the base so I’m not having it reinvent the wheel.
Should be fun to see what actually happens.
