Operational Failure Database
The definitive taxonomy of deterministic system collapse. Understand the symptoms, telemetry signals, and exact governance remediation protocols for the most critical agentic engineering failures.
Context Rot
Recursive Patching and Semantic Contamination
Context Rot occurs when an agent's context window becomes poisoned by its own failing outputs, leading to a death spiral of recursive patching where the agent attempts to fix bugs caused by its previous hallucinated fixes.
Retry Inflation
Unbounded API Burn Rate Expansion
Retry Inflation is the silent financial killer of agentic workflows. It happens when an agent, or a network of agents, gets caught in an automated loop of generating, failing tests, and retrying without a hard economic ceiling.
Tool Chain Recursion
Unconstrained MCP Access and Orchestration Chaos
Tool Chain Recursion occurs when an agent with global Model Context Protocol (MCP) access begins calling tools that call other tools, leading to server overload, unpredictable side-effects, and potential data exfiltration.
Verification Collapse
The Crushing Weight of Hallucination Debt
Verification Collapse happens when the velocity of agent-generated code significantly outpaces the human capacity to review it, resulting in a backlog of probabilistic code that senior engineers must manually debug and QA.
Repository Drift
Loss of Deterministic State Control
Repository Drift is the gradual decay of a codebase's architectural integrity caused by dozens of uncoordinated, probabilistic agent patches that bypass established design patterns and governance boundaries.
Why Claude Loses Context
Session Degradation & Context Poisoning
Claude loses context because its semantic window fills with stale assumptions, previous errors, and recursive patches. It "loses the plot" when not constrained by a bounded cognition engine.
Why Claude Rewrites Unrelated Files
Ghost Dependencies & Architectural Drift
Claude rewrites unrelated files due to "over-editing" hallucination. It infers ghost dependencies and mutates the repository outside of its authorized scope.
Why Claude Patch Loops Happen
Recursive Retries & Token Burn
Claude enters patch loops because it attempts to fix syntax errors without clearing its context of the broken state. It literally "starts patching its own patches" infinitely.
Why Cursor Spirals
Uncontrolled Execution & Context Collapse
Cursor spirals when the agent loses deterministic alignment with the codebase architecture and starts generating hallucinated implementations that break downstream logic.
Why Agentic Systems Fail in Production
Governance Theater & Probabilistic Variance
Agentic systems fail in production because organizations rely on prompt engineering (Governance Theater) instead of hardcoded deterministic runtime middleware. You cannot prompt an agent into safety.
The Unreliability Tax
Human Review Overhead That Erases AI Productivity Gains
The Unreliability Tax is the hidden cost layer that accumulates when organizations deploy probabilistic AI systems without verification gates. Every AI-generated output that requires human review, correction, or rollback adds a tax on the theoretical productivity gains. When the tax exceeds the savings, AI deployment becomes a net negative.
Synthetic Model Collapse
Training Data Contamination From AI-Generated Content
Synthetic Model Collapse occurs when AI models are fine-tuned or retrained on datasets contaminated by previous AI-generated content. Each generation of synthetic data introduces subtle distribution shifts that compound across training cycles, producing models that drift toward homogeneous, low-variance outputs that fail on edge cases.
AI Margin Collapse
The Point Where AI Feature Costs Exceed Revenue Contribution
AI Margin Collapse is the financial inflection point where the variable inference costs of AI-powered features exceed the marginal revenue those features generate. It occurs when organizations scale AI features without unit-level economic monitoring, and the growing token/compute costs silently erode gross margins below sustainable thresholds.
Shadow Agentic Execution
Unsanctioned AI Agent Deployments Operating Without Governance
Shadow Agentic Execution occurs when employees, teams, or departments deploy autonomous AI agents that take real actions (write to databases, send emails, modify codebases, interact with customers) without centralized governance, audit trails, or kill switches. These shadow agents operate in production environments without the organization knowing what permissions they hold or what actions they execute.
MCP Tool Sprawl
Uncontrolled Model Context Protocol Server Proliferation
MCP Tool Sprawl is the governance failure that occurs when organizations adopt the Model Context Protocol without centralized control over which MCP servers are installed, what tools they expose, and what permissions those tools grant to LLM agents. With 17,000+ community MCP servers available and zero built-in authentication in the MCP specification, every installed server becomes an unaudited attack surface.