Pattern Automation
← Learn glossary

Learn

What is agent cost optimization?

Agent cost optimization balances model choice, context size, heartbeat frequency, and tool calls so roles deliver accepted work per dollar — not minimum tokens on a leaderboard.

In depth

Inference spend scales with window size, model tier, number of tool loops, and schedule frequency. Cost optimization targets cost per accepted task — the economic unit that matters to the business.

Levers on Neuro OS:

- Role-scoped budgets — caps per role per day/week (Budgets) - Model routing — smaller model for triage; frontier model for final draft only - Context discipline — RAG top-k, summarize history, skills in git instead of re-prompting procedures - Heartbeat tuning — don't poll every minute if hourly suffices - Receipts — identify roles that loop without shipping output

Cheap runs that humans reject are expensive. A slightly costlier run with 90% acceptance beats a penny run rewritten entirely.

Examples

  • Triage with small model; escalation run uses larger model once
  • Weekly finance brief instead of daily if anomalies are rare
  • Cache retrieved KB chunks for repeated ticket types
  • Budget alert when outbound role exceeds $50/day inference
  • Kill switch on runaway tool loop after N steps

Related terms

FAQ

Should I always use the cheapest model?

No — optimize accepted output cost, not token price alone.

What is a healthy budget for one role?

Start from pilot volume × observed tokens per accepted task; add 30% buffer.

Run governed agent roles on a company OS — not only definitions in a glossary.