Definition · AI agents

What is an AI agent?

The word "agent" is doing real work in that sentence. An agent has agency: a persistent objective and the latitude to choose actions toward it over time. That distinguishes it from a single model call (which produces one output for one input) and from a fixed automation (which runs the same predetermined steps every time). An agent is defined by four things — a goal, tools, memory, and a level of autonomy. Take any one away and you have something simpler.

The four components of an AI agent

1. Goals

An agent starts from an objective, not a single instruction — "keep every open ticket updated," "get all five status inputs collected by Thursday," "answer this question using our own docs." The agent's job is to figure out the steps that satisfy the goal and to keep working until it is met or it reports back that it cannot. A good agent also knows when a goal is done, when it is blocked, and when it should stop and ask.

2. Tools

Tools are how an agent acts on the world beyond generating text: searching, reading a database, calling an API, creating a ticket, sending a message. A model on its own can only produce words; tools are what turn a decision into an action. The set of tools an agent has — and the permissions on them — largely defines what it can actually do and how much damage a mistake could cause.

3. Memory

To pursue a goal over time, an agent has to remember: what it has already tried, what it is waiting on, what it learned two steps ago. This is state. Short-term memory holds the current task's context; longer-term memory lets an agent carry a commitment across days and pick up where it left off. Without memory, every step starts from zero and nothing that spans time can be held.

4. Autonomy

Autonomy is how much the agent decides and does without a human in the loop — and it is a spectrum, not a switch. Most useful agents sit deliberately below "fully autonomous," acting freely on low-stakes reads while gating high-stakes writes behind human approval.

• Level 0 · AssistantResponds to prompts. No standing goal, no action between turns — a chatbot.
• Level 1 · SuggestRecommends an action; a human carries it out. The agent proposes, you dispose.
• Level 2 · Act with approvalProposes and executes actions, but each consequential action needs a human "yes." Reads freely, writes only on approval.
• Level 3 · Supervised autonomyActs within set bounds and reports back. No per-action approval, but a defined scope and an audit trail.
• Level 4 · Full autonomySets sub-goals and acts without per-action oversight. Rare in production for anything that can cause harm.

Agent vs chatbot vs workflow

Three terms often get blurred. A clean way to hold them apart: a chatbot responds, a workflow repeats, and an agent decides. A chatbot answers the message in front of it. A workflow runs a fixed sequence of steps the same way each time. An agent holds a goal, chooses its own next step, uses tools, and adapts when reality doesn't match the plan — which is why an agent can handle the messy, branching work that a rigid workflow can't.

A worked example: a coordination agent

Definitions land better against a concrete case. Ayven — the product this site is about — is an AI agent whose goal is coordination: making sure work that people committed to actually gets followed through. It's a useful example precisely because it is not maximally autonomous; it shows what the four components look like when an agent is built to be trusted with real work.

• Goal: keep a commitment moving until it is genuinely done — a status collected, a ticket unblocked, an answer chased from every person who owes one.
• Tools: it reads and posts in Slack and Microsoft Teams, and reads and writes work in trackers like Jira, YouTrack, and Asana (plus GitHub for code-linked checks).
• Memory: it holds state as durable "missions" — what it asked, of whom, what it's still waiting on, what it already tried — so a chase can span days rather than resetting each morning.
• Autonomy: it sits at Level 2 to 3 — it acts on its own clock because a deadline says so (a temporal gate), but every write to a tracker is approval-gated, and every action is logged.

The part that matters for trust is how it defines "done." A coordination agent that graded its own homework would be worse than useless. So when work is linked to a ticket, Ayven checks the tracker rather than taking a chat message as proof:

Ayven APP The reply said "shipped." I checked the linked ticket instead of taking the word for it — it's closed and the PR is attached. So this counts as done, and the goal is met.
Done — Jira · verified

That is the whole shape of an agent in one screen: a goal (follow the commitment to done), a tool (the tracker), memory (it knew what it was waiting on), and bounded autonomy (it verified before it claimed, and a human approved the write). "Done" is scoped to what the ticket shows — not a promise, and not the agent's own say-so.

Common types of AI agents

"AI agent" covers a range of designs. A few common types, by what they're built to do:

• Task or workflow agents — carry a multi-step task to completion, calling tools along the way (e.g. a coordination agent, a research agent).
• Conversational agents — goal-directed assistants that hold a dialogue and take actions within it, beyond a plain chatbot.
• Tool-using / API agents — orchestrate external systems: booking, querying, updating records.
• Multi-agent systems — several specialised agents that collaborate, one delegating sub-goals to others.

What they share is the definition above: a standing goal, tools to act, memory to persist, and a chosen level of autonomy.

Frequently asked questions

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

A chatbot responds to the message in front of it and stops. An AI agent holds a standing goal, decides its own next step, uses tools to act, and keeps working across multiple steps or days until the goal is met or it reports it is blocked. Every agent can converse, but not every chatbot is an agent.

Are AI agents autonomous?

To varying degrees — autonomy is a spectrum. Some agents only suggest actions for a human to carry out; others act within set bounds and report back. Most agents built for real, consequential work deliberately keep a human in the loop for high-stakes actions, acting freely on low-stakes reads while gating writes behind approval.

What are examples of AI agents?

A research agent that searches sources and synthesises a cited answer; a coding agent that edits and runs code; a customer-support agent that resolves tickets end to end; and a coordination agent that chases commitments across a team and verifies completion in the tracker. Each pairs a goal with tools, memory, and a level of autonomy.