AgentCache forks slash multi-agent token bills

The interesting part here is not “multi-agent” so much as “cache discipline as architecture.” If provider prefix caching is the billing lever, then fork-based session reuse is the right primitive, not another framework that sprays fresh contexts everywhere.

• –The library turns cacheability into a first-class constraint: shared prefixes, frozen cache-relevant params, and explicit cache-break detection
• –Its reported numbers are strong enough to matter in practice, with examples in the repo showing large cached-token shares and meaningful wall-time reduction on parallel worker runs
• –Cache-safe compaction is a smart addition because long-lived agent sessions usually fail on transcript bloat before they fail on reasoning quality
• –The real tradeoff is fragility: prompt edits, tool schema changes, and model swaps can silently destroy hit rates, so the diagnostic layer is as important as the fork logic
• –This is most compelling for coordinator/worker systems, research swarms, and any agent DAG where repeated prefixes dominate the cost profile