Key result
5 repeated jobs found in raw agent logs, with no task list.
together they cover 85% of the agent's writes
Setup
- Input
- Raw agent logs from τ-bench retail, with no task list
- Measured
- The share of the agent's writes the jobs cover
The question
An agent looks like it does a thousand different things. Underneath, how many jobs is it really doing?
What we found
From the agent's raw logs, with no task list, AgentCompile found 5 repeated jobs. Together they cover 85% of the agent's writes.
A few jobs, done again and again, are most of the work.
Why it matters
If most of the work is a few jobs, a few compiled jobs change most of the calls. Your agent keeps the rest: everything new, unclear or unusual.
Limits
- One domain: retail customer service. Your agent's jobs will be its own; a pilot starts by showing you which repeat.