Okay, we had an ambitious goal for Tinker—an autonomous AI agent that will build Pixel. But, after consideration, we decided to go simple phase by phase approach.
Phase 1 - We introduced looping agent which is connected to OpenAI API and speaks gibberish every 5 second internval.
Phase 2 - It now looks at tasks directory for any open tasks and moves them to ongoing then to done.
Phase 3 - We gave it shell read access with capability to prompt user with a shell commands to run.
Phase 4 - We dockerized this, so there is no inadvertent mishaps.
PR: https://github.com/kndlt/tinker/pull/1
Currently, the agent is able to create persistent docker image and do some work on it like creating files and what not.
Next steps: I will think about that tomorrow. I’m logging off.
