Fifty Steps to Nowhere

An agent capped at 50 steps keeps hitting the cap on a task that needs about six tool calls. The transcript shows it re-running nearly the same search with slightly different arguments. Raising the cap to 200 quadruples the cost with the same failure rate. What is the correct read?

Repeated similar calls mean the results are being served from a stale cache, so disabling the result cache will let each search return fresh output and the run finish

The step limit is a termination backstop, not a capability knob: the tool is returning output the model cannot act on, so fix the tool's results and errors

The cap is still too low for this workload, so remove it and bound the run with a wall-clock timeout instead

The model is sampling too conservatively, so raising temperature will break it out of the repeated search pattern and let it move on