Understanding the Spectrum of Agency in AI Systems

Not every AI system should be fully autonomous. One of the most useful design decisions is choosing the right level of agency for the problem.

This page treats agency as a spectrum, from static outputs to collaborative multi-agent systems.

The Five Levels

Level	Name	What the system can do	Typical examples
0	No agency	Return fixed or tightly scripted outputs	FAQ bots, static classifiers behind a simple UI
1	Routed workflows	Choose between predefined branches	workflow bots, intake triage, template routing
2	Tool-using assistant	Call external tools in bounded ways	internal copilots, retrieval assistants, support agents
3	Autonomous agent	Plan and execute multi-step tasks	research agents, coding agents, automation agents
4	Multi-agent system	Coordinate several specialized agents	planner-worker systems, reviewer-executor systems

At this level, the system does not really decide. It returns fixed or narrowly constrained outputs.

Use this when:

The system can inspect an input and select a predefined path.

Use this when:

This is often the first level where an AI system becomes genuinely useful in business workflows.

The system can:

but within clearer boundaries than a full autonomous agent.

At this level, the system can plan, adapt, and continue over multiple steps with minimal intervention.

Use this when:

This is where pages like AI Agents and Agentic AI become essential.

This is useful when one role is not enough.

Examples:

This can improve specialization, but it also increases complexity and coordination cost.

Choose the lowest level of agency that can still solve the problem.

Questions to ask:

Higher agency usually means:

That tradeoff is the real point of the framework. Agency is not a badge of sophistication. It is an architectural choice.