Rethinking Performance Engineering for Agentic AI

This book is free.

Your agent passed every load test. It's timing out in production anyway.

The performance playbook that served us for twenty years assumed one thing: the system does roughly the same work for every request. Agentic AI breaks that assumption. An agent decides at runtime how many model calls, tool calls, and reasoning loops each request needs, and every dashboard you own was built for a world where that number never changed.

This is a practitioner's book about what replaces the old playbook, written by a performance architect with two decades of production experience:

• Latency budgets with designed degradation paths, replacing single SLO targets that describe nothing
• Bounded autonomy: the four config values that turn an unbounded worst case into a fifteen-second one
• A token SLO per trace, with a real CI gate from the author's own pipeline that blocked a deployment every traditional metric approved
• Production-grade observability: the LLM metrics that matter (tokens, cache hit rate, context bloat, loop counts), instrumented once via OpenTelemetry and monitored through Langfuse, Splunk, and Arize
• Evals as the quality corner's monitoring system, AI gateway and orchestrator standards, and a 90-day path from one team's practice to an enterprise standard
• A first-week plan: five days, one agent, an artifact you keep every day

No theory dressed as practice. Every chapter ends with something you can apply this week, and the whole book reads in two sittings.

For performance engineers onboarding to agentic AI, and architects making agent fleets production-grade.