In a mature A2A ecosystem, the same capability is offered by multiple agents at different price/quality/latency points. Picking the right one for the current request is a runtime decision — negotiation. The patterns are similar to ad-auction or service-mesh routing decisions.
Static preference + fallback
Configured: agent A primary, B fallback if A unavailable. Simple, predictable. Doesn't optimize for cost or quality variance.
Price-quality tradeoff
Each provider advertises: cost per call, expected latency, quality score (from past benchmarks). Caller picks based on current request value: high-value → expensive/high-quality, low-value → cheap. Same logic as ad-server bidding.
Multi-shot voting
Critical decisions: call N agents in parallel, vote on response. Wisdom of crowds; catches outliers. Expensive but reliable. Right for high-stakes choices.
Auction model
Caller broadcasts requirement. Provider agents bid (price, expected quality). Caller picks. Complex; needs broker infrastructure; not common yet in 2026 but emerging in mature ecosystems.
Feedback loops
Track per-provider outcomes. Drift in quality? Down-weight. Cost increases? Re-evaluate. Negotiation isn't static; ratings update based on observed performance.