← Back to the lab
When an agent fails, the model is rarely the cause
I tested what changes when the model gets rules, tools and checks around it. The same request dropped from 17.3 to 8.4 minutes, with the same model.
17.3 → 8.4 minsame request, after the harness
Hypothesis
A smaller model in a good harness beats a frontier model in a bad one.
Result
| Same request | Before | After |
|---|---|---|
| Raw-material re-reads | 26 calls | 0 |
| Duration | 17.3 min | 8.4 min |
| Tokens | 2.09 M | 1.26 M |
A long task that used to make 121 calls and deliver nothing now delivers in 71 calls and 5.8 minutes.
From the OrganIA lab (agent-builder).
● Lab Log
One note from the lab every two weeks.
Experiments, real numbers and what went wrong. No spam, leave anytime.