Post

I was expecting more novel ways for working with agents. Better remote control, innovative ways to create plans or steer the agents, maybe more optimized token usage. But none of the ones I tried seems to deliver anything like that. The difference seems to be in how well they can render diffs.

Unironically I find that asking the models to output Simple Technical English (i.e. stop yapping) has more impact than harness differences 🤣