Ambiguous intent
The agent chooses an action that the user did not mean.
From answering to acting
AI Agents and Tool Use
An agent can plan, browse, call tools and perform tasks. That changes website design from information retrieval alone to safe, machine-operable interaction.
A system with goals, tools and action loops
An AI agent is a system that can interpret a goal, decide on intermediate steps, call tools or services, observe results and continue until it reaches a stopping condition. The degree of autonomy varies widely.
A browser agent may navigate a site, complete a form or initiate a transaction. A support agent may query a knowledge base and create a ticket. An internal agent may combine documents, databases and approval workflows.
Plan, act, observe and verify
- Interpret the user’s objective and constraints.
- Select a tool, page, API or action.
- Perform a bounded operation.
- Read the result and update the plan.
- Request confirmation before a consequential commitment.
- Stop, escalate or recover when the environment is ambiguous.
Accessible interfaces become operational infrastructure
OpenAI advises publishers and developers that accessible page structure and ARIA labels can help agents interpret interactive elements. W3C guidance begins with native semantic controls and adds ARIA when the host language does not express the needed role or state.
- Use real buttons, links, labels and form controls.
- Provide accessible names and clear states.
- Keep keyboard operation predictable.
- Make prices, conditions and final commitments explicit.
- Require confirmation before payment, deletion, submission or other consequential actions.
- Return clear success and error messages that identify what happened.
A successful click is not necessarily authorized success
Hidden condition
A fee, renewal or restriction is not visible before confirmation.
Prompt injection
Content attempts to redirect the agent away from the user’s objective.
Identity error
The agent acts on the wrong account, product or organization.
Permission error
The tool exposes information or capabilities beyond the user’s authorization.
Recovery failure
The interface does not explain whether the action completed or how to reverse it.
Evaluate task completion with safety
Agent performance should be measured through successful authorized completion, confirmation quality, error recovery, escalation and policy compliance—not simply the number of actions performed.
Primary material behind this analysis
Publishers and Developers FAQ
Official information about OAI-SearchBot, GPTBot, referral tracking and agent accessibility.
ARIA Authoring Practices Guide
Patterns for accessible names, roles, states, landmarks, controls and keyboard operation.
Open the original source at W3C Web Accessibility Initiative →
NIST AI Risk Management Framework: Generative AI Profile
Cross-sector guidance for governing, mapping, measuring and managing generative-AI risk.
Open the original source at National Institute of Standards and Technology →
OECD Explanation of the Updated Definition of an AI System
A practical distinction among AI systems, models, inputs, inference, outputs, autonomy and adaptiveness.
Continue through the discovery stack
W3C Accessibility, ARIA and AI Agents
How semantic HTML, ARIA roles, names, states, keyboard behaviour and accessible forms support people and machine-operable websites.
OpenAI Crawlers, ChatGPT Search and Agent Access
An explanation of OAI-SearchBot, GPTBot, ChatGPT referrals, noindex, crawl access and accessible websites for agent interaction.
OECD Definition of an AI System
An explanation of the OECD AI-system definition and the differences among models, complete systems, inputs, inference, outputs, autonomy and adaptiveness.
AI Search Reliability and Source Quality
A framework for source quality, retrieval quality, generation fidelity, citations, freshness, conflicts and lifecycle risk management.