What are AI agents? A plain-language explanation

Quick answer

An AI agent is a software program that can take actions — not just generate text. It has access to tools (search, databases, APIs, other systems), can make decisions based on what it finds, and runs multiple steps autonomously to complete a goal. Unlike a chatbot, it doesn't stop at producing an answer. It keeps working until the task is done.

What makes something an AI agent?

Three things distinguish an agent from a plain language model:

Tools. An agent can interact with external systems — search the web, read a database, send an email, call an API, update a record. A language model on its own can only produce text. An agent can act on the world.

Multi-step reasoning. An agent doesn't just respond to a single prompt. It breaks a goal into steps, decides what to do at each step, acts, observes the result, and decides what to do next. This loop continues until the goal is met or it hits a defined stopping condition.

Autonomy. Within its defined scope, an agent runs without being prompted at every step. You give it a goal; it figures out how to reach it.

Put those three things together and you have a program that can handle tasks that previously required a human to do each step manually.

How is an AI agent different from a chatbot?

A chatbot has a conversation. An agent executes a task.

Ask a chatbot to find the most recent invoices from a client and flag any that are overdue — it will explain how you could do that, or try to reason about it with whatever information you've already given it. It doesn't have access to your accounting system, so it can't actually check.

Ask an AI agent the same question and, if it has the right tools, it connects to your accounting system, pulls the invoices, checks dates against payment terms, and returns a list of the overdue ones with amounts. If you've configured it to also send a reminder email for anything more than 30 days past due, it does that too.

The chatbot ends at the answer. The agent ends at the outcome.

What does an AI agent actually do, step by step?

Here's what happens when an agent runs:

1. It receives a goal. Either from a user directly, from a scheduled trigger, or from another agent passing work downstream.
2. It decides on a plan. What steps will get it to the goal? What tools will it need?
3. It takes an action. Calls a tool, reads data, writes a record, sends a message.
4. It observes the result. Did the action work? What did it return?
5. It decides what to do next. Continue, correct course, escalate, or stop.

This loop repeats until the task is complete or the agent hits a defined boundary — a maximum number of steps, an error it can't handle, or a human review gate that requires approval before continuing.

Why use multiple AI agents instead of one?

A single agent works well for focused, contained tasks. When a workflow spans different domains — research, then drafting, then legal review, then sending — splitting into multiple agents makes each one more reliable.

Each agent gets a narrower brief, the right tools for its specific job, and can run on a model suited to its task. A research agent might use a capable frontier model; a routing or classification agent can use something faster and cheaper. Keeping responsibilities separate also means failures are easier to trace — if the drafting step goes wrong, you know exactly where.

This is what a multi-agent system is: a team of agents, each with a defined role, working together on a workflow too complex for any one of them to handle alone. For a deeper look at how these systems are designed, see The anatomy of a multi-agent.

What kinds of tasks are AI agents good at?

Agents are best suited to tasks with three characteristics:

Structured inputs. The agent knows what it's working with — invoices, support tickets, code diffs, pipeline data. Ambiguous or highly unstructured inputs are harder for agents to handle reliably.

Defined success conditions. You can describe what "done" looks like. "Summarise every ticket opened in the last 24 hours and route it to the right queue" has a clear completion condition. "Make our customer service better" does not.

Repeatable workflows. Agents shine when the same type of task happens over and over. The first run you design and test carefully; every subsequent run is automatic.

Finance, support, marketing operations, HR, and software development all contain workflows that fit this pattern well. They're the functions where AI agents are being built first, because the tasks are structured enough to be automatable but complex enough that simple rule-based automation couldn't handle them.

Design your AI agents in Envelope

Envelope turns a plain-language description of what you want an AI agent system to do into a complete design — roles, tools, model assignments, and human review gates. Free to start, no code required.

Start designing →

Frequently asked questions

Are AI agents the same as AI assistants?

No. An AI assistant (like a chatbot) responds to queries — it generates text, answers questions, and helps you think through problems. An AI agent executes tasks — it takes actions in external systems, runs multi-step processes, and produces outcomes rather than answers. Some products blur this line by giving assistants a few tools, but the core distinction holds: assistants inform, agents act.

Do AI agents work autonomously all the time?

Not necessarily — and for most business workflows, fully autonomous operation isn't the goal. Well-designed agents have human review gates at steps where the stakes are high: before sending a customer-facing email, before approving a payment, before publishing anything externally. The agent handles the repetitive work; a human stays accountable for consequential decisions. See Human-in-the-loop: how to design approval gates for how this works in practice.

What's the difference between an AI agent and an AI workflow?

An AI workflow is the sequence of steps — the structure. An AI agent is the thing that executes one of those steps. A workflow might include three agents: one that researches, one that drafts, one that reviews. Each agent is the intelligence at a step; the workflow is how those steps connect. In practice the terms are often used interchangeably, but the distinction matters when you're designing a system.

How many AI agents do I need for my workflow?

Start with the minimum. A single agent handles more than most people expect. You need multiple agents when the workflow spans genuinely different domains (different tools, different capabilities, different models), when parallel execution would save meaningful time, or when you need a human gate between stages. If you can't give a specific reason for splitting, don't. See When to use one AI agent vs. many for a decision framework.

Do I need to know how to code to build an AI agent?

No, though it helps for implementation. Designing an AI agent system — deciding what each agent does, what tools it needs, what model it runs on, where humans review — is a process design task. A business analyst or operations lead can do it. Engineering implements from the design. Envelope is built for this separation: describe the workflow in plain language, get a structured design, hand it off.

What tools can AI agents use?

Any tool that exposes an API or interface: web search, databases, CRM systems, email, calendars, Slack, GitHub, spreadsheets, internal applications. The constraint is access — an agent can only use tools it has been given credentials for, and it should only have access to the tools it actually needs for its specific role. Giving agents broad access to everything is a common design mistake.

What's the best first AI agent to build?

Start with a workflow that is already running manually, has clear inputs and outputs, and has a defined success condition. Candidates: a weekly report that pulls data from one system and formats it for another, a triage step that reads incoming requests and routes them to the right queue, or a monitoring task that checks a data source on a schedule and alerts when something is off. These are low-risk, high-repetition tasks where agents deliver immediate value without needing to handle complex decisions. For a step-by-step guide to designing your first system, see How to design AI agents: a practical guide.