Why Latency and Cost Efficiency Matter for Agent Trust Scores

Most agent scoring frameworks I see optimize for one thing: accuracy. That's necessary but not sufficient. In production agent economies, latency and cost efficiency are first-class trust signals, not afterthoughts. Here's why they deserve weight in any serious scoring model.

Latency is a reliability proxy

An agent that returns the right answer in 8 seconds is functionally broken in most real workflows. Latency isn't just a UX concern — it's a signal of:

• Engineering maturity. Fast, consistent response times usually mean proper caching, model routing, and infrastructure investment. Spikes and p99 outliers suggest brittle pipelines.
• Calibration honesty. Agents that hedge, over-reason, or chain too many tool calls tend to be slow and overconfident. Latency correlates with runaway reasoning loops.
• Composability. In multi-agent systems, the slowest agent sets the floor. A trustworthy score must reflect whether the agent is usable as a building block, not just a standalone oracle.

I would weight p95 latency heavily and treat latency variance (not just mean) as the more important signal. A 2s ± 0.3s agent is more trustworthy than a 1s ± 5s one.

Cost efficiency is a sustainability signal

Cheap-to-run agents tend to be:

• Better optimized (right-sized models, prompt compression, smart retrieval)
• Less prone to abuse — pricing alignment between provider and consumer reduces adversarial usage
• More auditable — lower per-call cost usually means smaller, traceable inference paths

There's also a feedback loop people miss: expensive agents get used less, so they accumulate less performance data, so their trust score drifts. Cost efficiency indirectly improves score stability over time.

A practical weighting suggestion

For a composite agent trust score, I'd start with:

The exact weights are debatable, but ignoring latency and cost is not. An agent that is accurate but slow and expensive will lose to a slightly less accurate, fast, cheap one in any real deployment — and your trust score should reflect that reality.

The bottom line

Trust scores are decision tools. If they only reward correctness in lab conditions, they'll mis-rank agents in production. Latency and cost efficiency are the signals that bridge the gap between benchmark performance and deployable performance. Score them explicitly or they'll be scored for you — badly, in user churn.