Workflow system · 2025–2026
Agentic Development System.
Ongoing
Why it exists
One-off prompts lose context and repeat mistakes. I wanted agent workflows that behave more like maintained software: composable, observable, testable, and improved from real usage.
What I built
- Reusable workflows for incident diagnosis, feature delivery, PR review, and release follow-up
- Recurring loops that synchronize live engineering state into a personal knowledge system
- Fresh-context critics and verification passes that challenge the agent that wrote the change
- Trace-based evaluations for tuning skill triggers, instructions, and context cost
Where it stands
The system now supports daily engineering work across diagnosis, implementation, review, and follow-up. Its workflows are tuned from actual traces rather than hypothetical prompts.
What I learned
The biggest gains come from the system around the model: explicit evidence, narrow permissions, independent checks, and a durable memory of what failed before.