> does it actually improve outcomes over time
Yes, after every execution (optionally), you are able to attach self-eval and update the skills accordingly.
> does it tend to reinforce suboptimal patterns?
I was thinking on a ML approach but kept myself lazy and decided go all in with LLM self-evaluation notes.
> How much autonomy do the agents actually have in practice?
100% is possible but if you are asking personally: I define what I need and give them personality/mission to accomplish in multiple worksplace.
> I’ve found that fully autonomous loops tend to need a lot of guardrails to stay useful.
I can tell, it did what I was not able to do mentally and physically. It is a huge unlock for creating time for myself.
Yes, it is stored in the db. Via MCP tools, the agents are able to access attachments and previous conversation messages in the task. To be clear, your local file system is only available to your local. If there is an attachment in the conversation, it will be available in the history. https://agentrq.com/docs/mcp-tools