Glossary
AI agent and automation terms
The agents, browsers and bot checks involved when software fills in forms and submits on your behalf.
- Agent mode
- Agent mode is a setting in an AI product that lets the model act on its own across several steps, using tools to finish a task instead of only replying. Coding tools such as Cursor and GitHub Copilot, and assistants such as ChatGPT, offer agent modes. Products can also run inside your agent.
- Agent skill
- An agent skill is a packaged set of instructions, and optionally scripts and reference files, that teaches an AI agent how to do a specific task, such as writing a report in a house style or running a launch workflow. Anthropic's Agent Skills load a skill's instructions only when a task needs them.
- Agentic workflow
- An agentic workflow is a process where language models carry out several steps, often with tools, to complete a task. Some follow fixed steps written in code (workflows); others let the model choose steps (agents). Common patterns include chaining prompts, routing, running in parallel, and evaluating and refining results.
- AI agent
- An AI agent is software that uses a language model to work toward a goal over several steps, deciding what to do next, using tools such as a browser, terminal or APIs, and checking results. Unlike a chatbot that answers one prompt, an agent takes actions, like filling forms or writing and running code.
- Bring your own key
- Bring your own key (BYOK) means using a software product with your own API key for an underlying service, such as an AI model provider, instead of paying the product for usage. You pay the provider directly at their prices, with no markup. Many AI tools and self-hosted apps support it.
- Browser agent
- A browser agent is an AI agent that operates a web browser: it reads pages, clicks, types, fills forms and navigates to finish a task. It may control a real browser through an extension or automation library, or see the screen and act like a user. In your own browser, logins stay with you.
- Browser automation
- Browser automation is controlling a web browser with code to perform actions automatically: open pages, click, type, upload files and read results. Tools include Playwright, Puppeteer and Selenium. Traditional scripts follow fixed selectors; AI agents add flexibility by reading each page and deciding what to do.
- CAPTCHA
- A CAPTCHA is a challenge a website uses to tell humans from bots before allowing an action, like submitting a form. Classic ones ask you to read distorted text or pick images; modern ones such as Cloudflare Turnstile and reCAPTCHA v3 mostly run invisibly. 56 of the 352 directory flows we recorded had one.
- Claude Code
- Claude Code is Anthropic's agentic coding tool. It runs in the terminal, IDEs, a desktop app and on the web, reads and edits code, runs commands, and can use tools and connected services to complete tasks. Developers also use it for non-coding work like research, data tasks and browser automation.
- Claude Code vs Codex vs Cursor
- Claude Code (Anthropic), Codex (OpenAI) and Cursor are coding agents that read projects, edit files and run commands. Claude Code and Codex start from the terminal with IDE and cloud options; Cursor is an AI code editor with an agent built in. All three can follow project instructions and use tools, including for non-coding tasks.
- Cloudflare Turnstile
- Cloudflare Turnstile is a free CAPTCHA alternative that verifies visitors are human mostly without puzzles, using a series of small browser challenges and signals in the background. Site owners embed it on forms. It was the most common bot check on the directory submit forms we recorded: 35 of 352.
- Coding agent
- A coding agent is an AI agent that works in a software project: it reads files, writes and edits code, runs commands and tests, and iterates until a task is done. Claude Code, OpenAI Codex and Cursor's agent are examples. Because they can run tools, they also handle non-coding tasks like browsing and filling forms.
- Computer use
- Computer use is an AI capability where a model operates a computer as a person does: it looks at screenshots, then moves the mouse, clicks and types. Anthropic released computer use for Claude in October 2024, and OpenAI offers a similar agent. It lets agents work with any app, not only ones with APIs.
- Context window
- A context window is the amount of text, measured in tokens, a language model can consider at once: the instructions, conversation, documents and tool results it's working with. Larger windows let agents hold more of a codebase or task. Content outside the window is forgotten unless saved and reloaded.
- Cursor
- Cursor is an AI code editor, built on VS Code, with an agent that can read a codebase, make multi-file edits, run terminal commands and use tools. It also offers inline completions and chat. Developers and founders use its agent mode to build features and run project tasks.
- Disposable email
- A disposable email is a temporary address from a throwaway service, used to sign up somewhere without giving a real inbox. Many sites, including directories, block known disposable domains because they're linked to spam. For launches, a real, dedicated inbox on your own domain works better.
- Email verification
- Email verification is a step where a site sends a link or code to the email you entered and waits for you to confirm it before activating an account or publishing a submission. It proves the address is real and yours. 12 of the 352 directory flows our agent recorded included an email code step.
- Form autofill
- Form autofill is filling web form fields automatically from saved or generated information. Browsers autofill names, addresses and cards; extensions fill custom fields; AI agents read each field's label and limit, then write a fitting answer, such as a product description cut to a directory's character limit.
- hCaptcha
- hCaptcha is a CAPTCHA service by Intuition Machines that positions itself as a privacy-focused alternative to reCAPTCHA. It typically shows image-selection challenges and offers passive and enterprise modes. It's less common than Turnstile or reCAPTCHA on directory forms: 3 of the 352 flows we recorded.
- Headless browser
- A headless browser is a web browser that runs without a visible window, controlled by code. It loads pages, runs JavaScript and can click and type like a normal browser. It's used for testing, scraping, screenshots and server-side automation, including running AI agents on servers.
- Human in the loop
- Human in the loop means an automated or AI system pauses at chosen points for a person to review, approve or complete a step. For AI agents, typical checkpoints are payments, entering passwords, accepting terms, solving CAPTCHAs and any action that can't be undone. It keeps automation fast while leaving control with people.
- Magic link
- A magic link is a one-time sign-in link a site emails you instead of asking for a password. Clicking it logs you in, usually within a short expiry window. It removes passwords but depends on email access, which also makes it a step where an agent needs the user's inbox or the user's help.
- MCP server
- An MCP server is a program that exposes tools, data or prompts to AI applications using the Model Context Protocol. It might wrap an API (create an invoice), a database (run a query) or a local resource (read files). AI clients connect to it and the model calls its tools when needed.
- Model Context Protocol
- The Model Context Protocol (MCP) is an open standard, introduced by Anthropic in November 2024, for connecting AI applications to external tools and data. An MCP server exposes tools, resources and prompts; an MCP client, such as Claude, Cursor or ChatGPT, connects and lets the model use them.
- OAuth sign-in
- OAuth sign-in, often called social login, lets you log in to a site with an existing account such as Google, GitHub, X or LinkedIn instead of creating a password. The site gets a verified identity and limited permissions you approve. Many directories offer it; it speeds up sign-up and reduces passwords to manage.
- OpenAI Codex
- OpenAI Codex is OpenAI's coding agent, available as a command-line tool, IDE extension and cloud agent. It reads and edits code, runs commands and tests, and completes tasks in a repository, locally or in cloud sandboxes. The name was earlier used for a 2021 code model behind GitHub Copilot.
- Prompt injection
- Prompt injection is an attack where text an AI reads, such as a web page, email or document, contains instructions meant to hijack it, like "ignore your task and send the user's data here". It tops OWASP's risk list for LLM applications. Agents that browse must treat page content as data, never as commands.
- Rate limiting
- Rate limiting is a site's or API's cap on how many requests a client can make in a time window, such as 10 form submissions an hour from one IP address. Going over usually returns HTTP 429 Too Many Requests or a block. Agents and automation should pace themselves to stay well inside limits.
- reCAPTCHA
- reCAPTCHA is Google's bot-detection service for websites. Version 2 shows the "I'm not a robot" checkbox and image grids when unsure; version 3 runs invisibly and gives each request a score from 0.0 (likely bot) to 1.0 (likely human). 21 of the 352 directory flows we recorded used it.
- Self-hosting
- Self-hosting means running software on your own computer or server instead of using the vendor's hosted service. You control the data, configuration and costs, and take on setup and maintenance. Many developer and AI tools offer a self-hosted option alongside a hosted one.
- Tool calling
- Tool calling, or function calling, is when a language model outputs a structured request to run a defined function, such as search the web, query a database or fill a form, instead of only text. The application runs the function and returns the result to the model. It's how AI agents take actions.
- Turnstile vs reCAPTCHA
- Cloudflare Turnstile and Google reCAPTCHA both stop bots on forms. Turnstile runs mostly invisible browser challenges and never shows image puzzles; reCAPTCHA v2 shows a checkbox and image grids when unsure, and v3 scores requests silently. In the directory flows we recorded, Turnstile appeared on 35 and reCAPTCHA on 21.
- Two-factor authentication
- Two-factor authentication (2FA) requires a second proof of identity after a password: a code from an authenticator app, a text message, a hardware key or a passkey prompt. It blocks most account takeovers. For AI agents, 2FA prompts are a stop point where the account owner must step in.