Types of AI agents a business actually uses

Search for "types of AI agents" and you get two kinds of answer. One is the textbook list, which is correct and was written for robots and game-playing programs. The other is a long catalogue of product categories, with a new type for every department.

Neither helps much when you are deciding what to hand off at work. What matters there is simpler. What starts the agent, what can it touch, how long does the job last, and where does a person come in?

Sorted that way, a business meets six kinds of agent. Most real jobs mix two or three of them.

1. Assistants that answer

The simplest kind. You ask, it answers, the exchange ends. Summarise this thread. Explain this contract clause. Draft a reply. Nothing outside the conversation changes.

This is useful and it is also where most teams stop, because an assistant that only answers needs you to paste in the context every time. The answer is only as good as what you copied into the box.

The upgrade is an assistant that reads your tools instead of your clipboard. "How much did we invoice last month?" is answerable only if it can read billing. A good one says so when it cannot, and asks for the connection.

2. Agents that act in your tools

Here the agent does the steps. It updates the CRM record after a call, files the action items in the tracker, drafts one reminder per overdue invoice and sends them. The output is a changed system, not a paragraph.

This is the type that deserves the most scrutiny before you buy, because mistakes leave the room. The question to ask is who says yes. Cedric's rule is short. Reading is free, and anything that sends, changes a record or spends money waits for a person, with the exact change on the card. How approvals work has the full table.

3. Scheduled and triggered agents

Some work has a clock. The Monday numbers, the daily check for stuck items, the month-end reconciliation. A scheduled agent does the same job on the same cadence and posts the result where the team reads it.

A close cousin starts on an event instead of a time. A new deal, a paid invoice or a merged pull request kicks off the run.

What separates a good one from a noisy one is reporting by exception. A check that finds nothing should say so in one line. A check that finds three mismatches should be hard to skim past. More in recurring work and automations.

4. Workflows that wait

Some jobs span days and depend on other people. Collect availability from five interviewers, then book the loop. Send the client an intake form, then set up the project when it comes back.

An agent built for single messages forgets these. A workflow agent holds the state. It asks each person, credits each reply to the job when it arrives, nudges the quiet ones, and picks up again when the last answer lands, even if that is Monday. Cedric's runs are bounded, so a stuck one stops and tells you instead of looping. See workflows that wait.

5. Multi-step researchers

Research agents take a question with no single source. Build or buy? Which vendor? What did our competitors change this week? They search, read the sources, compare, and come back with a brief.

The thing to check is whether the brief shows its work. A useful one says what is known, what is contested and what it could not verify, with a source on every number. Cedric files a build-or-buy brief like that in Notion on a yes, and the competitor tracker names a source for every move it reports.

6. Meeting agents

The last type sits in the call. The basic version transcribes. The more useful version is the same agent that works in your channel. It answers a question out loud when you say its name, from live data. It also treats "we should send them the security pack" as work to queue, not a line in the notes.

Two things to check. The agent should join as a visible participant, and actions picked up in the room should wait for approval afterwards, not happen live. Both are how Cedric works. See Cedric in meetings.

And the textbook types

For completeness, the classic list comes from Russell and Norvig's textbook, "Artificial Intelligence: A Modern Approach". It sorts agents by how they decide. A simple reflex agent reacts to what it sees right now. A model-based reflex agent also keeps track of the parts of the world it cannot see. A goal-based agent picks actions that reach a goal. A utility-based agent weighs how good each outcome is. A learning agent improves its own performance over time. Wikipedia's summary is a fair starting point.

These describe how a program chooses. They say nothing about whether it can read your CRM or who approves what it sends, which is why the six working types above are more useful when you are buying.

Where Cedric fits

Cedric is one teammate that covers all six. He answers in Slack, Discord or Telegram, acts in the tools you connect, runs scheduled and triggered work, holds workflows open for days, researches with sources, and joins your calls.

The reason to want one agent rather than six is memory. A correction made in one job holds for every job and every teammate. If someone tells him hardware goes under equipment, the reconciliation, the weekly report and the next draft all follow it.