tbh im getting so exhausted by the sheer amount of "secretary work" needed just to keep coding agents on track rn.
my workflow in cursor basically just looks like this:
- manually editing .cursorrules so it stops hallucinating the wrong ui library for the 100th time
- forcing the agent to dump a SESSION_STATE.md before i go to sleep so it remembers wtf we did tomorrow
- copy pasting that exact same summary back into the prompt the next morning bc it inevitably gets amnesia
we were supposed to be automating our coding but instead we just made a new job for ourselves: glorified ai prompt babysitter.
maintaining state by writing raw text files feels like a massive step backwards. why isnt there a tool that just sits in the bg, watches your git diffs and tool calls, builds a silent state graph of the decisions and auto-injects it? is everyone here seriously just hand writing .md files to keep their agents from going off the rails?
[link] [comments]