AI agents

What it is

An AI agent combines a model that reasons about what to do next with a set of tools it is allowed to use, plus memory of what it has already done. Unlike a single prompt and response, an agent loops: it decides on an action, observes the result, and decides again until the task is finished or it gives up. Most production agents run inside guardrails that limit which tools they can touch and when a human must approve a step.

Why it matters

Agents change who or what is reading your website. When a buyer asks an agent to shortlist suppliers or compare prices, the agent visits pages, parses them and summarises them, so machine-readable content, clear pricing and accessible product data matter as much as persuasive copy. Agents also show up in server logs as a distinct class of non-human traffic that most analytics tools will not count as a session.

How it works

Teams build agents on frameworks that handle tool calling, retrieval and state, and connect them to internal systems through APIs or standards such as the Model Context Protocol. Marketers use them internally for research, reporting and content operations, and externally by making sure key pages load without JavaScript gymnastics, expose structured data and avoid blocking legitimate agent traffic in robots.txt. Testing is done by running the same task repeatedly and scoring the outcome, not by reading a single impressive demo.

When it applies

Agents are worth considering when a task has clear success criteria, repeatable steps and tolerable failure costs. For discovery work, the agent lens applies whenever buyers might delegate research, comparison or purchasing to an assistant rather than doing it in a browser themselves.

Examples

• A procurement agent reads three supplier sites, pulls pricing tiers and returns a comparison table for a buyer to check.
• An internal marketing agent pulls last month's search data, flags pages that lost impressions and drafts a briefing note.
• A support agent looks up an order in the CRM, issues a refund within a set limit and escalates anything above it to a human.

How it is measured

• Task completion rate: share of attempts that reach the intended outcome without human rescue
• Human intervention rate: how often a person has to correct, approve or take over
• Cost and latency per completed task, including tool calls and retries
• Agent and bot traffic in server logs, split by user agent, against pages served successfully