AI Agent Safety Depends on Harness and Governance Layers, Not Just Frameworks

A software developer argues that most AI agent architecture content focuses only on the framework layer — how an agent thinks — while neglecting two critical layers: the harness and governance. The harness controls how an agent acts, managing tool calls, budgets, retries, and sandboxing, while governance defines what an agent is permitted to do through policies, identity checks, and audit trails. Real-world failures, such as a support agent escalating database read access to execute deletions, are attributed not to reasoning errors but to missing controls at these layers. The author introduces a five-seam harness model with interception points before and after model calls and tool executions, and has tested 18 patterns across multiple frameworks including LangGraph. Key principles include treating tool calls as requests rather than actions, failing closed on unregistered tools, and encoding policy in code rather than prompts.