Properties of good agent tasks:
- scoped tasks, such that the agent is not fed a long context, or tons of code
- closed loop: the agent can evaluate its own work
- anything that’s not mission critical (dashboards, debugging tools, etc.)
- boring stuff or prototypes/experiments or things you’d never have time to try
- reproduction cases from user issues
- rubber duck to bounce ideas against in lieu of human
- you evaluate, take what’s reasonable and finalize.
The last point is the most imp. one.
READ THE FUCKING CODE!
Clankers are ass at coding as most of their code is learnt from the internet.
Slow the fuck down - Mario Zechner, creator of the pi coding agent
Friction builds understanding and taste. This is where you actually learn. All of this requires discipline and agency. All of this requires humans.
Maybe the real agent was me all along?