Is this related to AI agents?

Only by name. It is a semantic role from linguistics used in natural language processing.

Because the roles are prototypes defined by property clusters, not strict categories with clean boundaries.

In semantic role labeling, which supports information extraction, translation and question answering.

AI Glossary / Agents and automation / Advanced

Proto-agent

Definition

A proto-agent is one of two semantic proto-roles from linguistics: the participant in an event who acts, causes, moves or experiences, as opposed to the proto-patient, the participant who is affected. It is a cluster of typical properties, not a strict category.

What it means

Traditional grammar tried to give every event participant one crisp role: agent, patient, instrument and so on. Real language resists. In the sentence the committee approved the plan, is the committee a single agent or a collection? Proto-roles answer: do not force the choice.

The proto-agent is defined by a list of contributing properties: the participant causes the event, acts voluntarily, moves, exists independently of the event, and so on. The more properties a participant has, the more agent-like it is, and the more likely grammar treats it as the subject.

The proto-patient has the mirror list: being affected, changing state, not existing independently. Most real participants are mixtures, and that is the point of the theory: roles are matters of degree.

The idea, from linguist David Dowty's 1991 work, became important in computational linguistics because it gives machines a practical way to label sentence participants without resolving questions philosophers have argued about for decades.

Editorial note: this inventory files the term under agents and automation, but it is a language-analysis concept. Its connection to AI runs through natural language processing, not through workflow agents.

How it works

In language analysis, identifying who did what to whom is called semantic role labeling. Proto-roles simplify the task: instead of choosing among dozens of fine-grained roles, a system decides how agent-like and how patient-like each participant is.

The contributing properties act as cues. Volition, causation and movement push toward proto-agent; being affected and changing state push toward proto-patient. The subject of a typical sentence scores high on proto-agent properties.

Machine-learning systems learn these cues from annotated text. Features such as position in the sentence, the verb's voice and surrounding words predict how human annotators would label each participant.

The output feeds downstream tasks: information extraction, question answering and summarization all depend on correctly identifying who did what, especially in long or ambiguous sentences.

Three practical examples

The storm destroyed the fence: the storm scores high as proto-agent through causation without volition, which is why grammar still treats it as the subject.
She was given the files by the courier: the courier carries the proto-agent properties despite not being the sentence's grammatical subject.
The clerk stamped the form: in multilingual analysis, proto-roles let one scheme compare how different languages mark who acts and who is affected.

Think of it this way

Think of classifying birds. A checklist of features, feathers, beak, eggs, covers robins and penguins without demanding that every feature be present. Proto-roles do for event participants what the checklist does for birds.

The limit: the checklist classifies real birds. Proto-roles classify the meanings of words in sentences, so the features are interpretive judgments that annotators, and models trained on them, must apply consistently.

Worked example

extracting who did what from contracts

Suppose a firm wants to extract obligations from contracts: who must deliver, who must pay, by when. The sentences are long and the participants shift positions constantly.

A semantic-role approach labels each participant's proto-agent and proto-patient scores around every obligation verb. The party carrying the action properties attaches to the duty; the affected party attaches to the benefit.

Passive constructions and embedded clauses, the cases that break simple templates, are handled because the scheme relies on properties rather than word position. The scenario is invented; property-based labeling is the standard technique.

The extracted obligations still get human review, because legal consequences deserve more than a probabilistic who-did-what.

Illustrative case study

A fictional news-analytics company builds a system to record which organizations are reported as taking action. It treats grammatical subjects as actors and works until it meets passive headlines: thousands of layoffs announced, with the real actor buried or absent.

Retraining the extractor on proto-role labels, rather than subject position, recovers the acting organization in passive and complex sentences, and the absence of an actor becomes itself a recorded fact about the reporting.

The invented lesson: who-did-what is a property question, not a position question. Proto-roles exist because grammar alone will not answer it.

Common misconceptions and mistakes

A proto-agent is a simple kind of AI agent. It is a linguistics concept about event participants, unrelated to workflow agents.
The roles are strict boxes. They are clusters of properties; most real participants are partial members of both.
It only matters for grammar class. Who-did-what labeling underlies extraction, translation and question answering in practical NLP.

Business use

You will not deploy proto-roles directly, but systems you buy for contract analysis, media monitoring or review summarization rely on this layer of language analysis.

When evaluating such tools, test passive and complex sentences from your own documents: that is where who-did-what errors hide, and proto-role quality is what you are probing.

Check your understanding

An extraction tool records the wrong actor in passive headlines. Why does proto-role labeling handle this better than subject-position rules?

Answer: Proto-roles identify the actor by properties like causation and volition rather than grammatical position, so passive constructions still resolve who acted.

Related terms

Learn next

AI for Business Owners covers what to ask when buying document-analysis tools: https://aiforbusinessandlife.com/academy/ai-for-business-owners. It is an adoption course, not a linguistics course. [C2]

Sources

[S8] Jurafsky and Martin, Speech and Language Processing (3rd edition draft). Full URLs are in the source ledger. Verified 11 October 2026.

Previous: Protected class Next: PROTO - PATIENT