A dynamic planner selects and sequences actions at runtime to pursue a goal while adapting to new information. Unlike a static workflow, it may explore different strategies and produce a different action sequence on each execution. section 18.1
In agentic systems, dynamic planning makes authorization part of decision-making rather than a fixed barrier at an application boundary. Each proposed action is evaluated against its principal, action, resource, and current context before execution. section 18.0 section 18.1
A governed dynamic planner must account for:
- Runtime choice: the agent decides which action to take rather than following a predetermined sequence. section 18.1
- Tool-based execution: selected tools may reach sensitive data or external systems. section 18.1
- Adaptive behavior: new information can cause the agent to revise its plan. section 18.1
- Continuous authorization: each proposed action is checked under the conditions in which it occurs. section 18.1
A governed planning loop
Before invoking a tool or accessing data, an agent constructs a request describing its intended action. A policy engine evaluates the principal, action, resource, and context, then allows or constrains execution using the enterprise authorization architecture. section 18.0 section 18.1
