What is an AI agent? A practical definition

An AI agent is software that uses a model, approved tools, and a repeating decision loop to pursue a stated result. It reads the available context, chooses an action, observes what happened, and continues until the work meets its finish line, reaches a stop condition, or needs a person.

That definition is narrower than current marketing usage. A chat assistant may answer well without owning a job. A fixed automation may carry a reliable sequence without deciding what to do next. A person can exercise authority, judgment, accountability, and physical presence that software cannot.

An AI agent can replace human evidence gathering, reconciliation, drafting, scheduling, and record maintenance when the job and checks are explicit. People still keep the judgment, authority, relationships, and physical work the software cannot legitimately carry.

Publisher disclosure: FidelicAI publishes this definition and sells role-specific AI agents. The category definition does not prove that any particular agent is useful, safe, or worth its price.

Four routes can produce the same visible draft

The first output can look similar across chat, automation, an agent, and a person. The ownership pattern is different.

Anthropic defines agents as models that direct their own processes and tool use rather than following a fixed script. Its practical loop is plan, act, observe, adjust, and repeat. OpenAI’s agent guide identifies a model, tools, and instructions as core components and recommends agents for work where judgment over unstructured inputs makes fixed rules hard to maintain.

Both sources sell model or platform access. Their definitions clarify the architecture; they do not establish that a vendor’s finished role works in a buyer’s business.

The loop needs an exit before it needs more autonomy

An agent run needs a finish line and several ways to stop. It can finish with an accepted result, pause for an approval, report missing or conflicting evidence, hit a safe operating limit, or fail in a way that leaves the work recoverable.

The execution-loop guide follows those steps in detail. The short version is six questions:

1. Which business record is authoritative?
2. Which result must exist at the end?
3. Which tools may the agent read from or write to?
4. Which checks must pass before the result is accepted?
5. Which action requires a person or qualified professional?
6. Which condition stops the run and preserves what already happened?

Done when: a fresh reviewer can identify the source, permitted action, result, check, approval state, and stop reason without asking the agent to explain its own work.

A role label is not job ownership

Calling software a sales agent, finance agent, or AI employee does not transfer responsibility. The public role must state its ordinary work products, allowed records and connections, checking method, approval boundary, and response to a miss.

The anatomy of a fidelic agent applies that test to FidelicAI. A current role such as FARO, the AI SEO strategist, publishes its search audit, query-to-page map, correction queue, evidence sources, checking rules, connection set, rates, limit, and owner-approval boundary. Those records make the claim inspectable. They still do not promise rankings, traffic, or sales.

The agent versus chatbot guide is the better route when the decision is between direct conversation and continuing job ownership. The buyer guide is the better route when the job is clear and the vendor decision remains open.

More freedom is not automatically better

Autonomy is permission to choose and act within a boundary. It is useful only when the work needs that choice and the failure cost is controlled.

NIST’s AI Risk Management Framework separates governance, context, measurement, and risk management. It is voluntary guidance, not a product certification. The useful transfer is simple: identify who is accountable, measure the important result and failure, and choose controls from the actual use and consequence.

Use fixed software when the path is known and exceptions are rare. Use an agent when the input varies, the next useful step depends on what it finds, and the finish line can still be checked. Keep a person at decisions involving licensure, rights, money movement, binding commitments, unfamiliar high-stakes judgment, or a relationship that expects human accountability.

Test one representative job

Choose work that recurs and has a real consequence, but remains reversible during the test. Record the current human time, source records, accepted result, correction burden, approval point, and failure path. Run the same job through the proposed route.

Count the owner’s assignment, access, checking, correction, approval, filing, and repair time. A strong demonstration can still fail this test when the buyer carries the operating method afterward.

The finishability question supplies the acceptance record. The security page supplies the current architecture boundary. The work-product promise states the remedy for defined delivery misses without promising an outside business result.

Browse the current AI agent directory only after the job and finish line are clear. If no current role matches the work, use chat, automation, a person, or wait.