Agent vs Operator
Not every model-powered capability is an agent. The Fieldbook separates systems that can plan and direct their own path to a goal from fixed-spec task dispatchers, tools and sub-agents. Getting the label right changes how a system should be built, tested and governed.
Agent
A self-directing unit pointed at a goal, not a function. Carries context and exercises judgment under ambiguity.
- Pointed at a goal
- Decides what to do, when, and when it is done
- Reasons and adapts when it hits something off-spec
- Carries a persistent stance / role
Sub-agent
Local agency within a scoped lane, monitors signals, decides, and acts within its remit.
- Local judgment
- Owns one lane
- Decides within scope
Operator
A bounded task-doer: given input X, produce output Y. Dispatched, never self-directing.
- Bound to a function
- Dispatched: does not decide whether/when to run
- Breaks predictably on off-spec input (a feature)
- No stance: borrows agency from whoever dispatched it
Tool
A called capability with no agency of its own.
- Invoked
- No judgment
Gate
A human or policy checkpoint in the flow.
- Human / policy review
- Approval boundary
The classification lens: agency, not the presence of a model
Of 103 classified units, 82 hold real agency and 16 are operators. Almost none are marketed as operators.
Show more
The label and the behaviour do not line up. Whether that is marketing running ahead of the engineering, or a category still forming, is not something the corpus can settle yet. What it does change is governance: a system labelled loosely tends to be governed loosely.
What is the difference between AI automation and an AI agent?
Automation follows a sequence you defined in advance. An agent decides its own next step against a goal and can change course when something fails. If a plain script could do the same job without losing anything, it is automation, even with a model inside it.
What makes something a real AI agent?
Deciding what happens next. A real agent plans its own route to a goal, picks its own tools and recovers from a failed step. A product that always runs the same five steps is a workflow with a model attached, whatever it is called.
Does using an LLM make something an agent?
No. A fixed workflow can call a model for one classification or extraction step and still be a fixed workflow. Agency is about who chooses the next move, not about whether inference happens somewhere in the pipeline.
What is a sub-agent?
An agent handed a narrow job by another agent. The parent keeps the overall goal, delegates a scoped piece of work, and uses the result. It is the usual way to keep any one agent’s context, tools and permissions small.
Why does it matter whether something is called an agent?
Because it changes how you test and govern it. A fixed workflow can be tested exhaustively. An agent makes choices nobody listed in advance, so it needs monitoring, evaluation and hard limits on what it can reach. That ongoing oversight is the core of the NIST AI Risk Management Framework (NIST).