I spent a lot of time creating a way for agents to track their work, and keep going. A few days ago, I found Steve Yegge's ‘beads’ project.
(see here: https://github.com/steveyegge/beads).
It is good and solves the problem. It's fun watching all these new tools show up. Programming is new again.
interesting find, will give this a go!
i got into the habits of letting the coding agent write detailed git commits with bulletpoints. how do you use the beads output?
You basically just tell it to ask ‘bd’ for the next task and it handles dependencies and stuff. And then setup another prompt that’s good at creating tasks. Works surprisingly well.
I did something like this last summer where I gave the agents tools to manage their memories. There was a paper recently about achieving error-free long task sequences with checking agents and towers of hannoi memory that got amazing results.
