Every tool you trust was built to notice absence. The new failures are aggressively present.
Production AI agents now hold credentials, take real actions, and make judgment calls at speed. Nothing in the classic operations toolkit can see a system that is confidently wrong. Latency is fine. Error rate is fine. The service returned 200. The content was the failure.
AGENTS ON CALL is the day-two book. Shelves of good books teach you to build the thing; this one is about running it, written by a practitioner who built the discipline in production. By the last page you can put an AI system on call: objectives it can be held to, incidents it can be walked through, permissions with hard edges, and a bill that cannot quietly explode.
The book names the discipline — and hands you its words.
- semantic failure
- confidently wrong output with no error code — the signature failure of systems built on models
- the drawer
- your unaudited machine-credential inventory, now a hundred badges per human
- the action gate
- independent validation no eloquence can talk past — the actor never grades its own homework
- the trust ladder
- recommend → approve → act; the promotion path for any automation
- zero-model-call path
- a revenue-critical flow engineered to complete with no model at all
Written for the people day two lands on.
I.The engineer with the pager
Your company shipped an assistant, and the first strange transcript is climbing through support. Your runbook doesn’t cover software that can be wrong without being down.
II.The director setting policy
The board asked what the company’s newest software is allowed to touch — and who answers for it. Chapter 16 is a ninety-day plan you can hand a leadership team.
III.The security engineer
Machine identities outnumber your humans a hundred to one, and some of them now make decisions. This book makes the drawer boring again.
Sixteen chapters in four parts, and one map of the discipline.
The Collision
The new failure class · what survives from SRE · the five-plane map you can score your company on by Friday.
Running AI Reliably
Answer-quality SLOs and evals · observability for minds · incidents on transcripts · agents as identities · the language attack surface · orchestration · token FinOps · zero-model-call resilience · change management for nondeterminism.
Letting AI Run Production
The AI-SRE tool wave · guardrails for automated remediation · the on-call org when agents share the pager.
The Road
The ninety-day plan · three honest case studies · the last line the whole book earns.
Start free. Each essay is a chapter’s core idea, standing alone.
Vijaybhasker Pagidoju
A senior technology leader who has spent his career running production infrastructure at scale — release nights, incident bridges, the pager that goes off at 2:47. He wrote this book because nothing on the shelf covered the systems he was being asked to run. He lives in Missouri. This is his first book.
Launch readers get the ninety-day plan, free.
Sign up and Chapter 16’s ninety-day plan arrives the week the book ships — the page you can hand a leadership team. Chapters and essays as they land, launch news, nothing else.