Recommendations

Changes to suggest once the data shows where spend goes. Each entry states the change, what it saves, and the use case report that measured it. No figure without a report in ../00-docs/use-case-reports/.

Harness token configs (opt-in, they change agent behavior)

Candidates, not yet measured:

• shorter default replies (output is the largest share of spend)
• keep long sessions instead of restarting (cache writes cost ~40x a cached read)
• delegate only large tasks to subagents
• model per task (small model for search and trivial work)

Other candidates

• update or remove MCP servers that send agents into loops
• model per intent, from intent labels