Agent harness

As explained in Building Reliable AI Systems

A harness is orchestration infrastructure that manages an AI agent's loop, state, tools, approvals, tracing, and output channels.

An LLM that only returns text is not yet an agent. A harness coordinates tools, results, state, and recovery, turning tool integrations into a working agent loop. section 7.9 section 7.9.1

Separate orchestration from execution:

• The harness owns the agent loop, state, tool routing, approvals, tracing, and output channels. section 7.9.1
• The sandbox runs code and commands in isolation; credentials and sensitive data remain in the harness. If execution crashes, the harness can restore a checkpoint. section 7.9.1

Implementations

OpenAI’s agents SDK and Codex CLI illustrate model-native harnesses paired with sandboxed compute. OpenClaw illustrates a model-agnostic, community-extensible harness; NemoClaw uses OpenShell for enterprise sandboxing. section 7.9.1

Figure 7.4 The harness owns the agent loop, memory, and LLM communication. The sandbox executes code in isolation with no direct model access. Tasks flow right; results flow left. If the sandbox crashes, the harness recovers.

Use the diagram to trace task flow from harness to isolated sandbox and results back, including recovery after a sandbox crash. section 7.9.1

How Building Reliable AI Systems covers it

How deep it goes

Explains control-plane and compute-plane responsibilities, including isolation, credentials, and checkpoint recovery.

Examples

Compares OpenAI’s agents SDK and Codex CLI, OpenClaw, and NemoClaw with OpenShell.

What it assumes

Builds on MCP tools and contrasts harnesses with standalone LLM prompt-response behavior.

What you can do after

Evaluate agent architectures by distinguishing orchestration responsibilities from isolated code execution.

How it connects

Connects agent loops and tool routing with MCP tools, memory, approvals, tracing, and sandboxed execution.

liveBookmanning.com

Read Building Reliable AI Systems on liveBook Get the book at manning.com

Related concepts

• context engineeringhas part · Context engineering is the deliberate selection, organization, and management of information supplied to an AI system to support effective task performance.
• language modeluses · A language model is a computational model that represents patterns in linguistic tokens to estimate sequence probabilities and predict or generate language from context.
• AI agentpart of · An AI agent is a computational system that uses artificial intelligence to interpret information, select actions, and pursue goals through interaction with an environment.
• large language modeluses · A large language model is a machine-learning model, typically neural, trained on extensive text data to process and generate language.

This explainer was generated by Manning's knowledge graph from the text of Building Reliable AI Systems. The section links point to the parts of the book it draws on.