An AI agent is a system where a language model makes its own decisions about what steps to take and which tools to use to complete a task, instead of following a fixed sequence a developer wrote in advance. That’s the key distinction, and it’s not just marketing language — Anthropic’s own engineering team draws this exact line in their guide to building effective agents: workflows are systems where “LLMs and tools are orchestrated through predefined code paths,” while agents are systems where “LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks.”
What Actually Makes Something an “Agent” Instead of Just a Chatbot?
A plain chatbot answers a question and stops. An agent is given a goal, and it can take multiple actions in a loop — check something, act on what it finds, check the result, and decide what to do next — without a human manually approving every intermediate step. The three ingredients that typically define an agent are: access to tools (search, code execution, file access, APIs), the ability to observe the result of using a tool, and the autonomy to decide the next action based on that result rather than following a script.
How Is an Agent Different From a Workflow?
A workflow is predictable by design — a fixed pipeline like “summarize this document, then translate it, then format it as an email” where each step is hardcoded regardless of what the model finds along the way. An agent, by contrast, might decide mid-task that it needs to search for additional information before it can summarize accurately, or that a different tool entirely is a better fit for the request than the one it started with. This flexibility is also the trade-off: workflows are easier to test and predict, while agents can handle novel situations but are harder to fully control and can take more steps (and cost) than expected on an open-ended task.
What Do AI Agents Actually Get Used For Right Now?
Common real-world uses include coding agents that read a codebase, write a fix, run the tests, and iterate if the tests fail; research agents that search the web across multiple queries and synthesize findings; and customer support agents that look up an order, check a policy, and issue a resolution without a human routing each step. What these share is a task with a clear success condition but an unpredictable number of steps to get there — exactly the kind of task where a fixed workflow would need to anticipate every possible path in advance.
When Should You Use a Workflow Instead of an Agent?
Anthropic’s own guidance is that simpler is usually better: start with the most straightforward solution and only add agentic complexity when it demonstrably improves outcomes. A predictable, repeatable task with a small number of known steps is usually served better by a workflow — it’s cheaper, faster, and easier to debug when something goes wrong, since you know exactly which step failed. Reach for an agent when the task genuinely can’t be fully specified in advance.
Frequently Asked Questions
Is every chatbot with tool access an AI agent?
Not necessarily. A chatbot that calls one tool in response to one message and then stops is closer to a simple tool-augmented assistant than a full agent. The “agent” label fits better once the system can chain multiple tool calls together, react to what each result tells it, and keep going until it decides the goal is met, without a human approving each intermediate action.
Do AI agents work without any human oversight?
In practice, most production agent setups include some form of human checkpoint, especially for higher-stakes actions like sending money, deleting data, or publishing content — the agent proposes an action and a person approves it, or the agent operates within a narrow, pre-approved set of permissions. Fully unsupervised agents exist for lower-risk tasks, but oversight tends to scale with how costly a mistake would be.
What’s the difference between an AI agent and RAG?
They solve different problems and are often combined. Retrieval-augmented generation is a technique for giving a model access to specific external information at answer time; an agent is a broader system for autonomously deciding what actions to take. An agent might use RAG as one of several tools available to it — retrieving relevant documents is just one action it can choose among many.
For the original framing this article draws on, see Anthropic’s engineering guide “Building Effective Agents”. If you want to go deeper on two building blocks agents commonly rely on, our guides to retrieval-augmented generation (RAG) and context windows cover the mechanics underneath.
