Web Agent — AI Glossary

Reference details

Related terms

• Computer-Using Agent (CUA)
• Tool-Using Agent
• AI Agent

How it works

1. Page state is observed in some representation — accessibility tree, filtered DOM, or a screenshot — and that choice determines what the agent can perceive and how many tokens it costs.
2. Interactive elements are identified and labelled so the agent can refer to a target unambiguously rather than by guessing coordinates.
3. An action is chosen and executed through browser automation: click, type, scroll, navigate.
4. The page is re-observed after the action, because the result is only knowable from the new state — this loop is the whole mechanism.
5. Progress has to be judged against the goal each cycle, since nothing in the page tells the agent whether the task is advancing or looping.

Example: Why per-step reliability caps task length

An agent that completes each step correctly 95% of the time finishes a 10-step task 0.95¹⁰ ≈ 60% of the time, and a 30-step task 0.95³⁰ ≈ 21%. At 99% per step the 30-step task reaches 0.99³⁰ ≈ 74%. Nothing about the model's reasoning changed between those two rows — the compounding did. This is why long web tasks need checkpointing and verification rather than better prompting, and why reported step accuracy and reported task success are not the same claim.

Commonly confused with

Conflated with browser automation, which is the tool rather than the agent. A script drives a browser along a path a developer specified and breaks when the page changes. A web agent decides what to do next from what it observes, so it can handle a page nobody anticipated and will also take actions nobody intended. It is also not retrieval: fetching a page's text for a model to read is a read-only operation, while this clicks, types and submits — which is why it needs confirmation before consequential actions.

When to use it

Reach for it when:

• A task requires interacting with a site that has no API.
• The path varies enough that a fixed automation script would break.
• Consequential actions are gated behind confirmation rather than taken autonomously.

Reach for something else when:

• An API exists, which is cheaper, faster and deterministic.
• The path is stable, where a script is more reliable than an agent.
• Long multi-step tasks without checkpointing, where per-step reliability compounds against you.

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Web Agent — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path