Use gateway-level system-prompt caching plus session compaction to cut multi-agent latency and spend

TrueFoundry·medium signal

Manage context at the gateway layer rather than per-agent: cache stable system prompts so repeated agent calls don't re-pay for the prompt, and run session-level compaction at the gateway to compress shared history before fan-out. This centralizes token control across a fleet of agents instead of duplicating the same context-management logic in each worker.