Tinker - Phase 2 Plan

Today’s Day 8 of Sprited.

Yesterday, I bootstrapped a new project called Tinker (tentative name). Tinker is an AI agent which is fully autonomous with system level access. The idea is that it would be the virtual employee who’d be responsible for building company’s public facing AI Pixel.

Tinker currently spawns Docker container and it has shell access to that container. So it can create files, edit files and whatever you can do in a shell.

Today’s goal is to get to a place where Tinker can read Github Issues and submit PRs. I will start with a very simple “hello world tasks.“

---

Phase 2 Plan:

1. Issue Parsing

   • Tinker should monitor a designated repo.

   • If an issue with a specific format is opened (e.g. #TinkerTask: hello-world), it should parse the description and treat it as a mission spec.

2. Workspace Setup

   • Clone the repo into the Docker container (or mount it).

   • Create a new branch for the task.

3. Code Generation

   • Based on the task description, generate a small hello_world.py file or similar.

   • Include a comment referencing the GitHub issue for traceability.

4. Git Ops

   • Commit the changes.

   • Push the branch.

   • Open a pull request with a templated message: “This PR was auto-generated by Tinker in response to Issue #X.”

5. Logging & Feedback

   • Log each step to a file and post a summary comment in the original issue thread.

   • Make sure to include a safety check so it doesn’t run on every random issue.

---

Why Github? I want to be able to closely monitor what the agent is doing in a github PRs. Github also allows you to access control user by user, also allowing you to lock certain branches for force push.

How would this help in building Pixel? We are basically creating a virtual employee who is full autonomous. In some sense, it is an automation of myself. Anything I’m doing manually, I want to automate.

AI agents get stuck, don’t they? Yes, in my experience, it is not yet good at handling large code bases, and it needs frequent interruptions. That’s the problem space of Tinker. It is going to improve that process little by little until we reach an virtual engineer which can really do things autonomously even when I’m on a vacation.