I think a running agent periodically checks if a process it started has finished.

that shouldn't be what eats up tokens tho

gpt 6 sol already made a lot of progress with caches

i have a feeling unreal agent might have decided to release now rather than getting sherlocked

i just think its very risky right now to spend too much time building harnesses or anything on top of codex or claude simply because frontier labs will just absorb whatever works

I assume the bigger problem is the cache expiring while you wait for the tool call to complete. A bunch of hosted providers only keep the cache alive for ~5 minutes - if you sit there waiting for a 5 minute tool call, you get to pay to reload the entire context into cache

"More tool work per model turn" could reduce the number of cache reads (or even cache misses) and associated cost?