Mastering AI Agent Architectures: The Optimized Core: Inference, Context, Caching, and Scheduling Engineering: 6 (Mastering AI Agent Architectures — The Styles Series) - Tapa blanda

Sinopsis

Were the verdicts about price, or about architecture? This book finds out.

Books 2–5 found that almost every multi-agent style beats a single agent on accuracy and loses on cost. This book puts a tuned inference core under all 65 styles of the series and switches on, one at a time, the techniques inference engineering offers — then re-runs every style.

• Cost anatomy: every bill split into calls, input tokens, output tokens and roles
• Prompt and KV caching — about a fifth off with no answer changed
• Semantic caching — safe only when tight, and dangerous inside a pipeline
• Batching and offline batch APIs — a clock decision
• Model cascades — the biggest lever, and why verifiers must never be cascaded
• Context compression with measured loss curves
• Draft–verify execution, and why it needs a cache to be cheaper
• Stragglers, hedging, placement and deadline-first scheduling

The tuned decision cards show which verdicts moved — monitoring and the 5× cap — and which did not, because they were structural all along.

Runnable code, frozen and tuned results under three locks at tag gauntlet-b6. Book 6 of 7 in Mastering AI Agent Architectures — The Styles Series.

"Sinopsis" puede pertenecer a otra edición de este libro.