An AI agent that can writes its own tools and sub-agents
retained text
Hello. I have been working on building a personal AI assistant called MAVIS that can write, debug and test its own tools and sub-agents if the user task cannot be solved by existing ones. The purpose is to have an assistant that can handle tedious tasks on my personal computer without me having to figure out the tools and workflows for it. I have a rough implementation now and would like feedback around what could be done better. Additionally, I also wanted to experiment on some ideas regarding the technical design so the architecture may be a bit awkward. Edit: MAVIS has its own harness (ONI) that monitors OS and network calls. Additionally, it has domain specific memory for different components that is maintained by async background workers. It also has a rudimentary pipeline caching mechanism along with a local streamlit dashboard where you can see stats such as cache hits, latency and token usage.
document fetched 2026-09-15 00:20 · canonical https://reddit.com/r/AI_Agents/comments/1wfxot5/an_ai_agent_that_can_writes_its_own_tools_and