3 ms·
What is the interplay between such delegation and token cache timeout? If Fable is truly active for 4 days, enough to keep the cache hot, the cost would be ast
by frumiousirc 6d ago
What is the interplay between such delegation and token cache timeout? If Fable is truly active for 4 days, enough to keep the cache hot, the cost would be astronomical. OTOH, if Fable idles while subagents are active, then each awakening is a cache miss.
- FearNotDaniel 6d agoTrue, that’s why I didn’t have a single Fable thread running for the whole time. Each phase had its own orchestrator that ran about 4-5h, in Fable or Opus depending on complexity, delegating to sub agents along the way - they had tasks broken down in a granular enough way that each returned in anything from 5 to 30 mins approx, then the lead agent did some validating and status updates before delegating the next step. So the main agent’s cache never went cold. Each phase raised its own PR and had an opportunity for human validation/code review before the next phase kicked off in a new context.