Cost-latency-reliability frontier for agentic workflows
A concept note extending the cost-latency-reliability frontier work (ZAN-RN-010) to explicitly cover test-time compute strategies for agentic workflows.
Whether test-time scaling strategies (verification, rollout selection, reflection) move the frontier in ways that differ meaningfully from the retry/routing frontier already measured for compound AI systems.
- Test-time compute frontier chart
- Comparison against retry-only baseline
- Verifier and rollout cost accounting
Run the same agentic task through a retry-only policy and a verifier-plus-rollout-selection policy at matched compute budgets, then compare resulting frontiers.
Test-time compute strategies produce a distinct, generally more favourable, frontier region than naive retries because verification and rollout selection can raise reliability without proportionally raising cost.
Test-time compute as a distinct frontier dimension
ZAN-RN-010 measured cost, latency, and reliability across routing and retry configurations. This note scopes whether test-time compute strategies, verification and rollout selection specifically, deserve separate frontier treatment rather than being folded into generic retry overhead.
Related research
View allCost-latency-reliability frontiers for compound AI systems
How should teams reason about tradeoffs between model quality, retries, routing, cost, and latency in compound AI systems?
A measurement model for AI deployment readiness
How should teams measure whether an AI system is ready for controlled production deployment?
Trace semantics for tool-using agents
What trace semantics are needed to explain failures in tool-using AI agents?