liveBook · Manning

AI governance

As explained in Context Engineering

Governance comprises the structures, processes, policies, and controls that direct how AI systems are developed and operated.

AI-specific governance is the set of structures, processes, policies, and guardrails that direct how AI systems are designed, developed, deployed, and operated. It turns organizational objectives and legal and ethical requirements into controls for model selection, data use, evaluation, monitoring, documentation, and incident response. section 9.1

A practical governance framework rests on six pillars: privacy, safety and security, accountability, transparency, fairness, and human oversight. These pillars are complementary: a weakness in one can compromise the wider system. section 9.1

Governance policies become engineering controls at several boundaries:

  • Context assembly: Use authorization and data minimization to determine which records, memories, and tool outputs are eligible. Apply redaction, retention limits, and trace masking to protect sensitive information. section 9.5
  • Execution: Treat tool outputs and other untrusted spans separately from authoritative instructions. Enforce least-privilege execution, output validation, and guardrails against unauthorized actions and data leakage. section 9.5
  • Oversight: Escalate consequential or low-confidence decisions through review queues and override mechanisms. Record the evidence needed to reconstruct disputed decisions. section 9.5 section 9.1.3
  • Operations: Version prompt templates, schemas, and retrieval settings as software assets. Validate changes through automated CI gates and use gradual rollouts. section 9.5

Governing a support assistant

A refund tool can execute autonomously below a small threshold, require human approval above it, and block execution when policy evidence is missing. To reconstruct a disputed recommendation, retain the request ID, authorization context, prompt version, model identifier, retrieved document IDs, tool arguments and results, and reviewer actions. section 9.3.6 section 9.1.3

For coding agents, govern both context access and actions. Classify sensitive data before it enters prompts, logs, memory, or external APIs, and provide only what the task needs. The agent's credentials—not its prompt—set what it can read and modify. Separate permissions for exploration, code changes, and higher-impact actions such as merging or releasing. section 11.4.3

No single framework fits every deployment. The OECD AI Principles provide principle-setting guidance; NIST AI RMF 1.0 structures lifecycle risk management; ISO/IEC 42001 specifies an auditable management system; and the EU AI Act establishes a risk-based legal framework. section 9.1

Policies alone do not secure a system. Longer contexts can make privacy instructions less reliable, while poisoned long-term memory can affect future agent behavior. Governance therefore needs enforcement at context and tool boundaries, not just instructions in a prompt. section 12.2.4 section 9.3.6

How Context Engineering covers it

How deep it goes
Defines six governance pillars and details controls for context access, tool execution, audit trails, oversight, and deployment.
Examples
Uses a customer support refund assistant and coding agents to illustrate auditability, approval thresholds, and scoped permissions.
What you can do after
Operate and configure agent access, approval thresholds, and audit traces to enforce governance policies.
How it connects
Connects governance to context engineering, prompt injection, human-in-the-loop oversight, and context as code.

Also explained in

Compare how 5 books explain AI governance →

Related concepts

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