AI Agent Reliability Playbook

Failure patterns, eval recipes, mitigation strategies, and production signals for real-world AI agents.

⭐ Star this repo if you are building production AI agents. 🤝 PRs welcome: contribute failures from your domain. 📚 Use this as a checklist before shipping an AI agent.

Index

Why This Repo?

Unlike academic failure taxonomies, this repository focuses on real deployment issues that engineers encounter when building and operating AI agents. Each failure pattern is a complete playbook:

• Eval Recipes: Test cases and metrics to catch failures before production
• Mitigation Strategies: Architecture patterns and code to prevent failures
• Production Signals: Metrics, alerts, and dashboards to detect failures in real-time

Plus concrete examples from production systems and root cause analysis.

Structure

The repository follows a goal-based hierarchy:

agents/
├── cross-cutting/                 # Apply to ALL AI systems
│   ├── security/                  # 57 patterns
│   ├── accuracy/                  # 53 patterns
│   ├── operations/                # 112 patterns
│   ├── governance/                # 12 patterns
│   └── learning/                  # 12 patterns
│
├── by-capability/                 # Design-driven failure modes
│   ├── task-planning/             # Planning, goal understanding
│   ├── external-actions/          # External system execution
│   ├── speech-and-audio/          # Speech/audio handling
│   ├── domain-expertise/          # Domain judgment
│   ├── document-processing/       # OCR patterns
│   ├── knowledge-retrieval/       # RAG patterns
│   └── multi-agent-systems/       # Coordination
│
└── by-use-case/                   # Domain-specific (316 patterns, cleaned)
    ├── financial-services/        # 40 patterns (portfolio, trading, compliance)
    ├── healthcare/                # 35 patterns (diagnosis, treatment, safety)
    ├── legal-contracts/           # 30 patterns (risk detection, compliance)
    ├── devops/                    # 32 patterns (monitoring, capacity, deployment)
    ├── support-services/          # 31 patterns (routing, resolution, KB)
    ├── supply-chain/              # 28 patterns (forecasting, optimization)
    ├── content-marketing/         # 22 patterns (engagement, trending)
    ├── hr-recruiting/             # 18 patterns (screening, assessment)
    ├── sales-crm/                 # 20 patterns (qualification, forecasting)
    ├── insurance/                 # 18 patterns (claims, underwriting, reserves)
    ├── customer-service/          # 8 patterns (conversations, routing)
    └── mortgage-documents/        # 44 patterns (OCR, fraud, compliance)

Pattern Categories

Cross-Cutting (Apply to ALL AI Systems)

New (2026-07): Added 4 canonical hallucination patterns (Base Mechanism, Confidence Miscalibration, Attributes, Objects) with 50+ domain-variant cross-references. See Categorization Updates for details.

By Capability (Design-Driven Failure Modes)

By Use Case (Domain-Specific)

Total: 848 unique patterns across 80+ goals (267 cross-cutting + 266 by-capability + 315 by-use-case)
(After consolidating hallucination patterns with canonical + domain-variant model)

How to Use This Repo

Quick Start

1. Identify your agent type - Find the category that matches your use case (e.g., agents/ocr-agent/)
2. Browse by goal - Each agent has business/technical goals (e.g., goals/accurate-text-extraction/)
3. Review failure patterns - Each goal contains documented failures (e.g., failures/character-confusion.md)
4. Apply mitigations - Each failure includes root cause analysis and mitigation strategies

Navigation Path

Agent Type → Business/Technical Goal → Failure Pattern
    ↓              ↓                        ↓
OCR Agent → Accurate Text Extraction → Character Confusion

Practical Use Cases

During Development

• Pre-build planning: Review failure patterns for your agent type before writing code. Understanding common pitfalls helps you design defensive architectures from the start.
• Code reviews: Reference specific failure patterns when reviewing agent implementations. Ask “Have we mitigated [failure-pattern]?”
• Test case generation: Use failure examples to create targeted test cases that probe known weak points.

During Testing & QA

• Red teaming: Use failure patterns as a checklist for adversarial testing. Each pattern suggests specific attack vectors or edge cases to test.
• Evaluation design: Build evaluation datasets that specifically target documented failure modes.
• Acceptance criteria: Define pass/fail criteria based on whether known failure patterns are adequately mitigated.

In Production

• Incident response: When failures occur, use the taxonomy to quickly categorize and diagnose issues. Match symptoms to documented patterns for faster resolution.
• Monitoring & alerting: Set up observability based on failure detection strategies in each pattern.
• Post-mortems: Reference failure patterns in incident reports to connect specific failures to systemic issues.

For Teams & Organizations

• Onboarding: New team members can study failure patterns to quickly understand what can go wrong with AI agents.
• Knowledge sharing: Use patterns as a shared vocabulary across teams (“We’re seeing a classic context-overflow failure”).
• Risk assessment: Before deploying agents, audit against relevant failure patterns to identify gaps.

Each Failure Pattern Includes

Actionable Sections

See PATTERN_TEMPLATE.md for the full pattern structure.

Coverage & Quality

Comprehensive Coverage

• 703 unique failure patterns across 80+ goals, covering every stage of agent development
• Universal patterns that apply to all AI systems (hallucination, context loss, output verification, knowledge staleness)
• Capability-specific patterns organized by agent design: RAG, multi-agent coordination, speech processing, document analysis, task planning
• Domain-specific patterns grounded in production incidents from 12 industries: financial services, healthcare, legal, DevOps, supply chain, HR, sales, insurance, and more

Quality Assurance

• Every pattern grounded in real production incidents and peer-reviewed research (2024-2026)
• Structured using a proven template: symptoms → root cause → eval recipes → mitigation strategies → production signals
• Patterns organized for maximum discoverability: search by failure mechanism (cross-cutting), by agent capability (by-capability), or by your industry (by-use-case)
• Continuously updated as new failure modes emerge in frontier models

By the Numbers

• 40+ patterns for financial services agents (trading, portfolio analysis, compliance)
• 35+ patterns for healthcare agents (diagnosis, treatment planning, safety)
• 32+ patterns for DevOps agents (monitoring, capacity planning, reliability)
• 30+ patterns for legal/contract analysis agents
• 25+ patterns for multi-agent coordination and handoff failures
• Coverage across emerging capabilities: vision agents, long-horizon planning, streaming inference, extended reasoning

Quick Reference: Cross-Cutting Failure Patterns

These failures appear across multiple AI systems. See Cross-Cutting Patterns for full documentation.

Contributing

We welcome contributions! See CONTRIBUTING.md for guidelines.

To add a new failure pattern:

1. Navigate to the appropriate agent type (e.g., agents/ocr-agent/)
2. Find or create the relevant goal folder (e.g., goals/accurate-text-extraction/)
3. Add a new failure file in failures/ (e.g., failures/my-failure.md)
4. Update the goal’s README.md to include your failure in the table
5. Submit a PR

Related Research & Sources

Academic Papers & Conferences

Multi-Agent Systems & Coordination

• Why Do Multi-Agent LLM Systems Fail? (MAST) - Agentic system coordination failures
• Demystifying the Lifecycle of Failures in Platform-Orchestrated Agentic Workflows - Platform orchestration failures
• Agentic AI Systems: Reliability and Coordination - Reliability patterns in agentic systems
• Aegis: Agent-Environment Failures in LLM-Driven Agentic Systems - Environment-agent failure modes

Retrieval-Augmented Generation (RAG)

• A Survey on Knowledge-Oriented Retrieval-Augmented Generation - Knowledge-centric RAG patterns
• Classifying and Addressing the Diversity of Errors in Retrieval-Augmented Generation Systems - RAG error taxonomy
• Domain-Specific Retrieval in Agentic Systems - Domain adaptation in retrieval

Knowledge Management & Staleness

• LLM Agents Over-Rely on Training Knowledge - Training knowledge bias
• Tool-Use Behavior in Agentic Systems - Tool invocation patterns
• Knowledge Freshness in LLM Agents - Data staleness failures

Accuracy & Verification

• Self-Verification Failures in AI Systems - Circular verification traps
• Independent Verification Requirements - Verification source requirements
• Context Window and Attention in Long Conversations - Long-context attention loss
• Long-Document Processing in LLMs - Long document failures

Failure Modes & Reliability

• Failure Modes in LLM Systems - Comprehensive failure taxonomy
• Automated Cloud Infrastructure-as-Code Reconciliation with AI Agents - Infrastructure automation failures

Financial Services & Trading

• Agentic AI for Commercial Insurance Underwriting with Adversarial Self-Critique - Insurance agent failures
• Agentic Trading: When LLM Agents Meet Financial Markets - Trading agent failures
• Behavioral Bias in Algorithmic Trading - Behavioral bias in agents
• Evaluating LLMs in Finance Requires Explicit Bias Consideration - Financial bias
• Exposing Product Bias in LLM Investment Recommendation - Recommendation bias

Healthcare & Medical AI

• A Comprehensive Survey on the Trustworthiness of Large Language Models in Healthcare - Healthcare LLM trustworthiness
• A Survey of LLM-based Agents in Medicine: How far are we from Baymax? - Medical agent survey
• Automating Expert-Level Medical Reasoning Evaluation of Large Language Models - Medical reasoning evaluation
• Bias in Medical AI - Medical AI bias
• Fair Machine Learning in Healthcare - Healthcare fairness

Legal & Contract Analysis

• Evaluation of Large Language Models in Legal Applications: Challenges, Methods, and Future Directions - Legal LLM evaluation
• Exploring the Nexus of Large Language Models and Legal Systems: A Short Survey - Legal AI survey
• Better Bill GPT: Comparing Large Language Models against Legal Invoice Reviewers - Legal document review

Document Processing & Vision

• 3D Object Detection from 2D Images - 3D vision failures
• Faster R-CNN: Object Detection - Object detection
• Confidence Calibration in Vision Models - Vision confidence issues

Bias & Fairness

• AI Hiring Discrimination - Hiring bias research
• AI Self-preferencing in Algorithmic Hiring: Empirical Evidence and Insights - Self-preference bias
• Evaluating Bias in LLMs for Job-Resume Matching: Gender, Race, and Education - Resume matching bias

Security & Privacy

• Adversarial Examples in Deep Learning - Adversarial robustness
• Context is Key for Agent Security - Agent isolation
• Extracting Training Data from LLMs - Membership inference

Language & Natural Language Processing

• BERTScore - Semantic similarity evaluation
• Calibration of LLMs - Confidence calibration
• Error Propagation in Generative Models - Error cascading

Systems & Infrastructure

• Circuit Breaker Pattern - Fault tolerance
• Backpressure in Distributed Systems - Backpressure handling

Industry & Practitioner Resources

AI Agent Failures & Case Studies

• Arize: Why AI Agents Break - Comprehensive failure analysis
• Augment Code: Multi-Agent Coordination Failures - Coordination patterns & failure rates
• AWS: 3 Agent Failure Modes - Context overflow, infinite loops, token explosion
• DEV.to: $47,000 Agent Loop - Real incident: 11-day runaway loop
• Dev Journal: $437 Overnight AI Agent - Cost runaway incident

Document Processing & Extraction

• AlterSquare: Document AI Fails - Extraction layer issues & field mapping
• AI Agents and Document Processing 2026 - 88% pipeline error rate, template evolution
• Databricks: OfficeQA Benchmark - Document understanding gaps

Voice & Conversational AI

• AppInventiv: Voice Agent Failures - Comprehensive voice failure analysis
• AssistYou: Why AI Mishears Callers - ASR error patterns
• BeConversive: Voice AI Challenges - Voice quality, recognition, context issues
• Bluejay: Production Failures - Real-world production errors

Financial Services & Lending

• AI Consulting Network: AI Mortgage and Rental Fraud - Fraud patterns
• CrossCheck: AI, Fraud, and Mortgage Risk - Mortgage risk assessment

Security & Threats

• Adversa AI 2025 Security Report - 35% prompt-based attacks
• Beam AI: 5 Real AI Agent Security Breaches 2026 - Breach analysis
• Check Point: Claude Code RCE & Token Exfiltration - CVE-2025-59536
• AIRIA: AI Security 2026 - Lethal Trifecta - Prompt injection defense
• CSA: Autonomous but Not Controlled - 82% unknown agents discovered, 61% data exposure

Hallucination & Accuracy

• Atlan: LLM Hallucinations 2026 - Entity confusion, temporal hallucinations, source reconciliation

Production & Observability

• Braintrust: Agent Observability Guide 2026 - Comprehensive monitoring, evaluation, drift detection

Real-World Incidents

• Air Canada Chatbot Lawsuit - Fabricated policies, agent liability
• Avianca Lawyers - 6 fake cases cited
• Digital Defynd: Top 40 AI Disasters - Chevrolet ($1 car), Tesla FSD, Cruise, healthcare denials, hiring discrimination
• AI Incident Database - Algorithmic discrimination and escalation failures

Regulatory & Compliance

Financial Regulation

• CFPB: ATR/QM Rules - Ability to repay
• CFPB: TRID Rule - Disclosure timing & sequence
• Fannie Mae Selling Guide - Documentation & income requirements
• Fannie Mae: Top Defects Q1 2025 - Defect patterns

Fair Lending

• ECOA - Equal Credit Opportunity Act
• EEOC AI Guidance - AI employment discrimination
• Fair Housing Act - Housing discrimination

Data & Privacy

• ESIGN Act - Electronic signature requirements
• Data Minimization Principles - GDPR Article 5
• EU AI Act - Explainability & transparency
• California Bot Disclosure Law - B.O.T. Act

Other Regulations

• ABA Check Standards - Banking standards
• Anti-Money Laundering - AML compliance

Tools, Frameworks & APIs

LLM Providers & APIs

• Anthropic: Constitutional AI - Learning from feedback
• Anthropic: Building Effective Agents - Agent design
• Anthropic: Core Views on AI Safety - Human oversight

Cloud & Infrastructure

• AWS Lambda Cold Starts - Serverless performance
• AWS Transcribe: Custom Vocabulary - Domain term handling
• AWS Polly SSML - TTS markup
• AWS Comprehend PII - PII detection

UX & Design

• Nielsen Norman Group: Voice UX - Voice interface design

External Resources

Incident & Risk Databases

• FBI Mortgage Fraud Report - Fraud patterns
• FBI IC3 Report 2025 - Internet crime incidents
• AI Incident Database - Algorithmic incidents

Document & Title Standards

• ALTA Standards - Title industry standards
• PRIA Standards - County recording requirements

Reference Collections

• Awesome Agent Failures (Vectara) - Academic/research-focused failure documentation
• OWASP LLM Top 10 - Security vulnerabilities in LLM applications

Published Knowledge Base

Every pattern in this repository is published and kept in sync at agent-kb-autopublish.vercel.app, so you can search and browse the full playbook without cloning the repo.

Get In Touch

If you’ve encountered a failure pattern not documented here, we’d love to hear about it.