Playbook
Actionable guides, operating policies, and step-by-step testing procedures — all informed by documented real-world AI failures. Stop reading about what goes wrong. Start implementing what works.
New here? Follow these four steps to go from zero to a tested, safe AI agent deployment.
Take the 3-minute quiz to discover your highest-risk areas.
Take the Quiz →Walk through 67 testing checkpoints before you ship.
Get the Checklist →Study real documented incidents so you don’t repeat them.
Browse Failures →Implement prevention with step-by-step playbook guides below.
Jump to Guides →In-depth guides from InspectAgents, built on lessons from real-world AI incidents.
End-to-end testing framework: hallucination detection, prompt injection, security validation, and production monitoring.
Deep technical analysis of the prompt injection attack, timeline, and exactly how to prevent it in your own chatbot.
Continuously updated database of AI agent failures, hallucinations, and security breaches. Learn what went wrong — and why.
Reusable operating contracts that define evidence boundaries, verification rules, and fail-closed behavior for your AI agents. Curated from the AI Agents Playbook by Tamar Peretz.
Evidence boundary policy: world-claims require authoritative sources. No speculation, no hedging, no hallucination.
Restrict evidence to user-provided artifacts only. No browsing, no training data claims. Maximum control.
Operating contract for public facts via web sources with mandatory inline citations.
Architecture + best-practices + regression-minded review gate as a normative operating contract.
Prevent overclaims and enforce terminology consistency. Catch subtle misrepresentations before they ship.
Require a numeric confidence line on every response so users know how much to trust the output.
Step-by-step procedures for fact-checking, quality gates, and verification workflows. Copy-paste-ready prompt templates included.
A repeatable evidence-check loop to run before producing any non-trivial output. Claim-by-claim verification workflow.
Define allowed sources and refusal conditions for your AI agent. Fail-closed by default — artifacts-only or authoritative sources.
Architecture, best-practices, and regression-minded review gate. Validate outputs before they reach users.
Structured self-check loop before final output. The model verifies its own claims step-by-step before responding.
Reusable daily-driver prompt template with evidence boundaries, output contracts, and fallback rules baked in.
How to request web browsing and produce verifiable, citation-grade references in agent outputs.
Technical articles on trust boundaries, attack surfaces, and control-plane failures in tool-using AI agent systems.
Where the first LLM-to-system boundary appears, and why exposure starts before you even have "agents."
Why multi-step orchestration increases risk, and what to audit in plan → tools → retry loops.
Context hub, truncation/ordering, tool loops, observability, and R1–R8 audit checkpoints.
Concrete checkpoints for auditing trust boundaries in tool-using, multi-step AI agent systems.
Session binding, memory reuse, routing/selection, tool enforcement, and observability failure patterns.
Operating contracts that set evidence boundaries, citation requirements, and fail-closed behavior. Your AI agent's constitution.
Step-by-step guides with copy-paste prompt templates. Follow the procedure, get a tested output.
The 67-point checklist and verification gates ensure nothing slips through before deployment.
Our failures database shows exactly what goes wrong when these practices are skipped. Learn from others' mistakes.
Open-source playbooks and references from the AI safety community.
Public, versioned knowledge base for building reliable AI agents. Policies, prompt templates, how-to procedures, and security articles.
by Tamar Peretz • andyagentlab.com
Source repo for the AI Agents Playbook. Browse raw policy files, prompt templates (.system.txt / .user.txt), and contribute.
github.com/Tamarper63
Start with the quiz to identify your biggest risks, then follow the playbook to fix them — all before your users find them first.
Most teams can't — find out in 2 minutes
Real AI failures analyzed • Free 2-minute assessment