AI agent glossary
Short answers first, then what to check. Definitions come from vendor or standards documentation where we link one, and from our own guides otherwise. Checked 3 October 2026. We have not tested any product named here.
AI agent
Software that takes actions for you with tools you allow, instead of only answering questions.
A chatbot replies with text and you do the rest. An agent can use connected tools such as email, a calendar, files or a code editor, and report back. Many products mix both, so ask what it can do without you and what needs your approval.
See: What is an AI agent?
Approval (or confirmation)
A step where the agent stops and asks you before it acts.
Vendors describe approvals differently. Lucas, for example, says it cannot take irreversible actions such as sending email or creating events without your explicit confirmation. Other vendors give no detail. Read the vendor wording, then test with a low-stakes task.
Token
A small piece of text that a model reads or writes, and the unit many AI products count and bill.
OpenAI says a token can be as short as one character or as long as a word, and gives a rough English rule of thumb of about 4 characters, or 100 tokens for 75 words. Counts vary by model.
Credits
A balance a product spends when you use it, instead of a flat monthly price.
How credits are spent differs by product and can change. Check how fast your own usage would drain them before treating a low plan price as the real cost.
MCP (Model Context Protocol)
An open standard for connecting AI applications to outside data, tools and workflows.
The MCP project compares it to a USB-C port for AI applications. In practice it is how an agent can reach things like your calendar or a database through a standard connector. Each connector still needs its own permissions check.
Integration (connector)
A link between an agent and another service, such as Gmail or a calendar.
Connecting gives the agent access to that account. Connect one at a time, note what it can read versus change, and know how to disconnect. Some vendors say disconnecting does not delete data they already kept.
Sandbox
A contained environment where an agent runs code or commands so it cannot touch everything on your machine.
Not every tool is sandboxed. Anthropic says Claude Code mods are not sandboxed and run with the same access to your machine as Claude Code, so install them only from sources you trust.
See: Claude Code mods, Anthropic · OpenAI Agents API explained
Reverse trial
A trial that starts you on paid features and then steps you down to a free tier unless you subscribe.
Lucas says new users get a seven-day reverse trial. It is different from a free trial that ends in a charge, so read what happens on day eight.
See: Lucas explained
Outcome-based pricing
Paying per result, such as a resolved conversation, instead of per seat or per message.
Intercom says Fin is priced per outcome. What counts as an outcome is the vendor definition, so check it before you estimate cost.
On-device (local) AI
AI that runs on your own computer or phone rather than a vendor server.
Underdog says its models run on your Mac and your content is not sent to a cloud AI service. That is the vendor statement, and the product also keeps an account and optional usage counts. On-device does not remove the need to check what it can access.
See: Underdog record
Invite-only / waitlist
Access is limited to people with an invite code or a place on a list.
Several new agents are early access. A listed price may not be open to you yet, and plan details can change by launch.
See: Free AI agents
Where to go next
New to agents? Start with What is an AI agent?. Ready to choose? Browse the directory, filter by published price, or compare by job to be done. How we label evidence is in the methodology.