AI agents are great at figuring things out — but some tasks shouldn’t require figuring out. Restarting services, resetting test data, deploying a build — these should work the same way every time.
When something is deterministic, make it a script. Don’t let the agent improvise what should be repeatable. A script gives you the same result every run, costs zero tokens to execute, and removes a whole class of “the agent did something slightly different this time” surprises.
“Script” is broad here — a bash script, a Makefile target, an npm run command, a Gradle task. You can use whatever your ecosystem already has. Agents already know how to call make, npm run, and ./gradlew — using familiar tools means less explaining.
Every step the agent re-solves from scratch is a step that can drift. npm run reset-dev doesn’t drift.
💡 Try this: Think of one multi-step task you keep describing to the agent. Make it a script instead — then tell the agent to use that.
How do you handle repeatable tasks with your agent?
🟢 Scripts and Makefiles — the agent runs them
🟡 I have some scripts, but the agent doesn’t use them
đź”´ I mostly just explain it to the agent each time
⚪ Haven’t thought about it
What’s a task you’ve scripted (or should script)?