Tinker on Day 13

PublishedJune 8, 2025

Over the weekend, I did some refactor on the reasoning loop on Tinker. There was so much technical debt built on top by the Claude Code and I needed to do a holistic review of the codebase and ask Tinker to clean things up.

Tip*: One thing that helped is to ask the coding agents to “I’ll give you 100 iterations, do a thorough research on …“ This helped the agents to look into nooks and crannies of the codebase and do a holistic overview.*

One of the things we noticed was that Tinker was using strict think → act → observe → decide loop and this I believe is less flexible than agent being able to plan their action own their own. I decided to simplify this into one node.

To make that happen, we are now using LangGraph’s create_react_agent and @tools attribute to greatly simplify the codebase.

Next Step is to introduce memory summarization feature to reduce the context window usage.

#sprited #tinker

Comments

Join the discussion

No comments yet. Be the first to comment.

Tinker

Part 9 of 9

Tinker is our internal virtual engineer with containerized shell access and self-improvement capabilities. It's main job will be to build and maintain Pixel. https://github.com/kndlt/tinker

Start from the beginning

Tinker - Autonomous Engineer of Sprited

Okay, I’m currently engineering the Pixel’s memory system and to be blunt, it’s not that fun. It was fun to imagine the architecture and all, but I really want to get back to the Machi project. So, naturally, I’m wondering: can I just create an AI ag...