Read the transcript
Every time I hand a task to an AI agent, the same questions come back. What needs to be clear before it starts? What context does the agent need? What can it decide on its own? And what evidence do I need before I accept the result?
For months, I collected and tested practices that answered them. Forty practices, each with a check for your own work. Together, they form the Agentic Engineering Periodic System.
Top to bottom, each row is a loop with its own clock. One agent run: minutes. One change: days. The setup around your agents: weeks. Team decisions: months. The rows are not levels.
Left to right, every row runs the same loop. Understand the code. Decide what to build. Give the agent its context. Hand the work over. Check what comes back. Find out why it failed. Inside a security boundary. Then spread what works, and start the next round.
Take one card: Curated context. Which sources does this task need? Did the agent actually read them? Did it avoid rebuilding what already exists?
This is the map I use in my own work. It has releases, and you can send me corrections. Start small. One task. One practice. One check. The system is free, at ase.academy/ps.
Rows: four loops
Task, Change, System, Organisation. Each loop has its own clock, from minutes to months. The rows are not levels.
Columns: one loop
Understand the code, decide what to build, give the agent its context, hand the work over, check what comes back, find out why it failed, inside a security boundary, then spread what works.
Each card: one check
Every practice comes with a test you can check on your own work.