Fresh agent sessions would miss steps I’d already corrected in earlier work. I could get an agent to follow the workflow during one session, then have to give it the same nudge again in the next.
I built a process that lets agents examine their work, record what they learn, and revise their own operational guidance.
Turning feedback into instructions
The agent reviews a cycle of work, looking at what it did and the nudges I gave it along the way. It decides what to retain from those experiences, including corrections and changes in direction.
It keeps those lessons in an Obsidian-compatible Markdown knowledge base. Periodically, it uses them to revise its operating instructions. Those changes take effect automatically, without a separate review step.
The revised guidance informs the next cycle of work and is available to subsequent sessions. I can also read the lessons in Obsidian.
Remembering SQLx cache preparation
One recurring omission was SQLx cache preparation. SQLx’s cargo sqlx prepare command saves the query metadata needed to build with its query macros in offline mode. Fresh agent sessions would neglect that preparation step, and I’d have to remind them.
The system picked up on the omission and updated its own guidance to include the step. Later sessions read that guidance and included the preparation in their work. I stopped having to repeat the reminder.
Using it in practice
I’ve used the system in my own work and with teams. The experience has been more streamlined and consistently accurate: agents carry the lessons forward, and I spend less time repeating the same corrections.
I use a separate workflow for keeping software documentation current.