What are AI agents, and which jobs should a small team give them?

What an AI agent is

Anthropic's guide Building effective agents (December 2024) defines the two shapes of AI system it sees in practice. Workflows are systems where models and tools are orchestrated through predefined code paths. Agents are systems where the model dynamically directs its own process and tool use. Under the hood, an agent is a loop: the model chooses an action, a tool runs, the result goes back to the model, and it chooses again.

Tools are what make the loop useful. Through a standard such as the Model Context Protocol, an agent can read a help center, file a ticket or update a record. The surrounding setup that decides which tools exist, what the agent sees and when it stops is its harness. Anthropic's write-up on long-running agents shows that agents drift or declare a job done too early when that setup is missing.

The same guide's advice is to find the simplest solution that works and add complexity only when needed. Many jobs need one well-written prompt with the right documents, not an agent.

Why a small team should care

• You already have agents. Claude Code, Codex, Cursor's agent mode and ChatGPT's agent features are agents with your permissions.
• The value is in boring work. Copying details from emails into a CRM, checking 40 product pages, gathering facts for a brief.
• The risk is in the tools. An agent that can send, pay or delete will eventually do it at the wrong moment unless a person approves those actions.

A worked example

Illustration: a made-up business. Willow Events, a nine-person event agency, lists its recurring tasks and scores each on how often it happens, how long it takes, how clear the rules are and how bad a mistake would be. "Answer client emails" scores high on time and high on risk. "Collect venue details from supplier emails into the venue sheet" scores high on time and clarity and low on risk.

The team gives the venue task to an agent: read new supplier emails, fill the sheet's twelve columns, flag anything missing, and never reply to anyone. An account manager checks the sheet each morning. After two weeks, the agent fills 140 rows; 9 needed a correction, all in the "capacity" column, which leads to one new rule in its instructions.

How to pick the first job for an agent

1. List ten tasks your team repeats at least weekly.
2. Score each on frequency, time, how clear the rules are, and the cost of a mistake.
3. Pick the one with high time and clarity and a low cost of mistakes.
4. Write its limits: the tools it may use, what it must never do, and who checks the output.
5. Run it on a few items, read every result, then widen it.

Common confusions

• "An agent replaces a person." In most small teams it replaces a dull step, and a person still checks the result.
• "More autonomy is better." Anthropic's advice is the opposite: the simplest system that does the job.
• "Agents don't need instructions." They need clearer ones than people do. A shared skill gives every agent on the team the same steps.

Questions

What is the difference between an AI agent and a chatbot?

A chatbot answers one message at a time and a person decides what happens next. An agent is given a goal, decides its own next step, calls tools (search, email, a spreadsheet, a CRM), looks at the result, and repeats until the job is done or a rule tells it to stop.

What is the difference between an agent and a workflow?

Anthropic's agent guide draws the line this way: in a workflow, models and tools follow predefined code paths; in an agent, the model directs its own process and tool use. Workflows are more predictable; agents are more flexible.

Which tasks should a small team give to an AI agent first?

Frequent tasks with clear rules, where a wrong result is cheap to catch and a person checks the output before it leaves the team: filling fields from emails, first-pass research, tidying records. Keep sending, paying and deleting behind a person's approval.