Harness Engineering: Design Patterns for Securing Long-Horizon Multi-Agent AI Systems
Build a secure control plane for agentic AI with design patterns for bounded autonomy, least privilege identity, trusted context, governed memory, and runtime recovery
Key Features:
- Design a security control plane for long running agents and delegated multi-agent work
- Contain prompt injection and memory poisoning through enforceable trust boundaries
- Use HarnessLab scenarios to connect design patterns with observable control decisions
Book Description:
Long-running AI agents need controls that hold across tool calls, delegated tasks, and persistent memory. Harness Engineering presents design patterns for securing long-horizon multi-agent AI systems through a control plane that governs autonomy, identity, intent, context, and memory. Learn to scope agent permissions, contain prompt injection, detect goal drift, and place human oversight where consequential actions require it.
Through architectural diagrams and HarnessLab scenarios, explore decision tracing, evaluation gates, runtime intervention, recovery, fleet coordination, and data governance. Build the engineering judgment to make agent behavior observable, authority accountable, and improvement measurable, so you can move from promising prototypes to controlled production operations.
What You Will Learn:
- Set autonomy budgets and risk based human approval gates
- Scope agent identities, credentials, and delegation chains
- Bind tool authority to intent and detect goal drift
- Contain prompt injection and govern persistent memory
- Use decision traces and evals to govern agent improvement
- Steer live runs and enforce fleet and data boundaries
Who this book is for:
For AI engineers, security architects, platform teams, and technical leaders designing or governing long running agentic AI systems. Researchers and graduate students will also benefit from its pattern based treatment of agent security and control. Familiarity with LLM applications, tool calling, access control, and distributed systems is recommended. The book supports readers moving from agent prototypes to accountable enterprise operations.
Table of Contents
- Harness as the Control Plane for Long-Horizon Multi-Agent Systems
- Autonomy: Graduated Control and Human Oversight
- Identity: Per-Agent Identity and Delegation Chains
- Intent: Contracts, Drift Detection, and Authorization by Purpose
- Context: Trust Tiers, Provenance, and Context-Poisoning Containment
- Memory: State Firewall, Ledger, and Forgetting]
- Observability, Evals, and Hill-Climbing Evidence
- Runtime Steering: Intervention, Policy Injection, and Safe Rollback
- Scalability: Fleet Coordination Without Cascading Failure
- Data Governance: Classification-Aware Tools, Egress, and Compliance
Customer Reviews
ThriftBooks ® and the ThriftBooks ® logo are registered trademarks of Thrift Books Global, LLC
