Where does agent waste actually come from: model calls or orchestration?
I’ve been digging into a problem I think agent observability still misses.
Most tools can tell us what an agent called, how many tokens it used, and what it cost.
But the more useful question is:
Why did those calls happen at all?
I started looking at this after AUDR came out and built a small open-source CLI called KORA Doctor to inspect agent traces.
Right now it looks for things like:
- repeated model calls
- missed cache or reuse
- deterministic work sent to an LLM
- expensive models doing simple work
- excessive planning and orchestration
The first useful feedback changed my thinking.
Someone running agents in production told me their biggest source of waste wasn’t repeated inference.
It was re-fetching the same tool data across multiple steps.
That means the waste can start before the model call. The orchestration layer can create unnecessary retrieval, tool calls, retries, and then more inference on top of that.
So I’m starting to think the bigger problem isn’t just “LLM cost optimization.”
It’s execution waste across the whole agent run.
For people running real agents:
Where do you see the most waste?
Model calls?
Tool calls?
Retries?
Retrieval?
Planning loops?
Repeated validation?
Something else?
I’m building KORA Doctor around what shows up in real traces, so production examples are especially useful.
KORA Doctor:
https://github.com/Krako-Labs/kora-doctor