A lot of architecture decisions are not bad decisions. They are just badly remembered.
That is the problem I wanted to explore in my latest build. Teams move fast, make the best call they can, and a few months later somebody asks why a database, platform, or tool was chosen. The answer is usually in Slack, in memory, or with the one person who happened to be in the room.
I built a small ADR agent with Google ADK to test whether that reasoning could be captured and reused more cleanly. One agent, a few tools, and a simple knowledge base. Nothing fancy yet, but enough to prove the architecture, the principles, and the RICE framing in code instead of leaving them as diagrams and theory.
What this first build does
- drafts ADRs
- looks up earlier decisions
- saves them into a simple knowledge base
- answers follow-up questions from what was already decided
This is also the step before the multi-agent system. I wanted to make the first useful single-agent version real before splitting the workflow across multiple agents.
Want the full walkthrough? Read the complete post at zulmehdi.blog.
Top comments (1)
The useful bit is making the reason inspectable before splitting the work across agents. ADRs rot when they become just a decision log. If the agent can keep the tradeoff, rejected option, and trigger for revisiting it together, the memory is much less ceremonial.