What are AI agents?

By Mark JonesUpdated 6 min read

The working definition

An AI agent is a model given a goal, a set of tools it can call, memory of what it has already done, and permission to keep going until the goal is met or a limit is hit.

Strip away the marketing and every agent is the same shape: perceive, decide, act, check, repeat. What separates a real agent from a demo is not the model. It is whether the loop is allowed to close without a human retyping the next instruction.

Agent, chatbot, automation: the three get confused constantly

Chatbot

A chatbot produces text in response to a prompt. It has no obligation to the outcome. When the window closes, nothing happened in the business.

Automation

An automation runs a path someone drew in advance. It is fast, cheap, and reliable right up to the first input that does not match the path. Then it stops or does the wrong thing quietly.

Agent

An agent is handed the outcome instead of the path. It picks the steps, calls the tools, notices when a step failed, and tries a different route. That flexibility is the whole value and also the whole risk, which is why scope and guardrails matter more than model choice.

Do we need a bot? Ask what it should do after you ask

Most bot questions resolve once you name the behavior the job needs. There are three, they overlap, and none of them is a maturity level.

Help me now

You ask in chat and get an answer or a draft in the same conversation. The product's obligation ends when you stop typing. Most assistants inside Slack, Teams, and Google Chat cover this on plans a business already pays for.

Take a job and bring it back

You hand over a task, the product works on its own for a while, and it returns with a result or a question. This is delegation, and the thing to check is whether the result was verified against your system of record or only described by the bot.

Keep doing agreed work

A routine runs on a schedule or when an event fires, without anyone asking each time. This is where identity matters: a recurring run often acts as whoever set it up, not as the person who benefits from it, so review what that account can reach.

Four honest outcomes

The evaluation is the same for all four: pick one job that runs at least weekly, write down how it gets done today, define what finished means, and run the product against that standard for a short period with the outputs checked directly. A custom build is one possible outcome, not the starting assumption.

The five parts of an agent that actually works

Where agents earn their keep today

Agents pay off where the work is repetitive in shape but variable in detail. That is the exact gap traditional automation could never cover.

Where agents fail, and why

Most failed agent projects fail for the same four reasons, and none of them are about the model being insufficiently smart.

How to deploy your first agent

Pick a workflow your team does at least twice a week, that has a clear finished state, and where a wrong answer is embarrassing rather than catastrophic. That last constraint is what makes the first deployment survivable.

Run it alongside the human for two weeks with every step logged. Compare outputs. Tighten the standing context where they disagree. Then remove the human from the middle and leave them on the approval step. That sequence gets an agent into production without a governance fight.

What the runtimes now ship

The major labs now sell managed agent runtimes, and controls that used to be bolted on, such as budgets, approval gates, memory stores, domain allowlists, and scheduled runs, increasingly ship from the provider. The specifics differ by vendor and change often, so check the provider's current documentation before designing around any one control.

The consequence for a buyer is that the loop itself is no longer the hard part. The work that still needs a builder is deciding scope, writing the standing context the agent runs on, wiring it into your systems of record, and setting the guardrails so the runtime's controls are actually used. That is what an AI operations partner does; the model provider gives you the engine.

FAQ
What is an AI agent?+

An AI agent is a model given a goal, a set of tools it can call, memory of what it has already done, and permission to run a loop until the goal is met. Unlike a chatbot, it owns an outcome rather than a response.

What is the difference between an AI agent and automation?+

An automation follows a path someone defined in advance and breaks when the input does not match. An agent is given the outcome instead of the path, so it selects steps, calls tools, and recovers when a step fails.

Are AI agents safe to run unattended?+

Reversible work can run unattended with step budgets, cost ceilings, and full logging. Irreversible actions such as sending, paying, deleting, or publishing should sit behind a human approval step.

How many agents does a business need?+

Start with one agent doing one job with a clear definition of done. Multi-agent orchestration is worth adding once a single agent is reliably in production, not before.

Do we need a custom agent, or will a product we already pay for do the job?+

Decide by behavior, not by vendor. If the job is an answer in chat, a delegated task brought back, or a routine on a schedule, first check whether a product on your existing plan does that as supplied or with configuration. Build only when the job depends on a system the product cannot reach.

Do AI agents lock you into one model provider?+

They should not. Keep the goal, tools, memory, and guardrails outside the provider so the model underneath can be swapped when the frontier moves.

Have a workflow that should be running without you?

Bring us your project