chapter eighteen

18 Authorization for AI agents

 

This chapter covers

  • Why authorization becomes the hardest problem in agentic AI systems
  • How AI agents interact with enterprise systems and their authorization requirements
  • Policy-aware agent loops and runtime authorization checks
  • Using policies to guide agent behavior, not simply to deny actions
  • Delegation and subagent authority in agentic systems

A year after launching the Customer Collaboration platform, ACME started experimenting with a new class of tools powered by artificial intelligence (AI). What began as a few internal prototypes quickly spread throughout the organization. Engineers used agents to analyze logs and troubleshoot deployments. Sales teams utilized them to prepare account briefings. Customer support tested agents that could summarize cases and suggest responses.

The early results were impressive. Tasks that once took hours of manual effort could now be done in minutes. The engineering team developed integrations that let agents access internal systems, retrieve documents, and carry out routine operational tasks. The company started to envision an AI-enabled workforce where software agents worked alongside employees.

To coordinate these efforts, ACME hired Trudy, a seasoned technology leader, as Senior Director for AI. Trudy’s mandate was simple in principle but ambitious in scope: help ACME deploy AI agents safely and effectively across the enterprise.

18.1 Why authorization is the hard problem in agentic AI

18.2 How agents interact with external systems

18.3 A policy-aware agent loop

18.3.1 Authorization as part of the feedback loop

18.3.2 Defining operational boundaries with policy

18.3.3 From predefined workflows to governed behavior

18.3.4 Practical implications

18.4 Beyond denial: guiding agent behavior with policy constraints

18.4.1 From reactive enforcement to constraint-aware planning

18.4.2 Integrating constraints into the planning loop

18.4.3 Reducing trial and error

18.4.4 Policy as a source of guidance

18.4.5 From control to collaboration

18.5 Childproofing the control plane

18.6 Delegation and subagents

18.6.1 Delegation as data

18.6.2 Enforcing delegation at runtime

18.6.3 Policies that enforce delegation

18.6.4 Allowed and denied behavior

18.6.5 Delegation as constrained authority

18.6.6 Extending the policy-aware architecture

18.7 Cross-domain delegation among agents

18.8 Summary