Add an agent-cost-reduction benchmark scenario

[madeinplutofabio]

Summary

NCP's adoption pitch is "reduce repeated orchestration cost in agentic systems". A targeted benchmark that compares an LLM-only agent loop against an LLM+NCP loop on the same workflow makes this pitch concrete and defensible.

Acceptance criteria

• Pick one repeatable agent workflow. The lead-qualification graph (Add a practical lead-qualification MCP example graph #29) is a natural fit.
• Run two scenarios on the same input dataset:
  • A) LLM agent does the full workflow through prompt/tool orchestration each call.
  • B) LLM agent calls one NCP graph (via ncp-mcp-server) that runs the deterministic part of the workflow.
• Measure at minimum:
  • Number of LLM calls per workflow invocation
  • Total prompt and output tokens
  • End-to-end latency
  • Traceability and debuggability (qualitative: what can you reproduce from the trace alone?)
• Document assumptions honestly (model version, prompt strategy, dataset, cost-per-token used for projections).
• Add results under bench/ or docs/. Include raw measurement data, not just summary numbers.

Why this matters

This benchmark is potentially very powerful if done carefully. It gives adopters a defensible cost-reduction number to take to their teams.

Out of scope

• Closed-source models that cannot be fully scripted.
• Implying NCP is a drop-in replacement for LLMs. The point is to show where deterministic graphs replace orchestration overhead.

Where to read

• BENCHMARK.md for the existing benchmark methodology to align with
• bench/datasets/ for the existing dataset format
• The lead-qualification graph issue (Add a practical lead-qualification MCP example graph #29) for the workflow under test