We keep adding sophisticated capabilities to bots — rules, language understanding, LLMs, memory, tools — until almost everything gets called an AI agent. But where does the bot actually stop and the agent begin?
Not whether it uses an LLM. Not whether it has tools. The more useful distinction:
A bot follows a path. An agent determines the path.
The bot: a known path
A support bot asked “Where is my order?” might run a fixed sequence — identify intent, look up the order, retrieve status, respond. An LLM can power any step, but the developer still decided what to query, what to check, and what to return.
One fixed sequence, run the same way every time.
That's the traditional strength of bots: predictability. Per OpenAI's framing, using an LLM without letting it control workflow execution doesn't make something an agent — agents use the model to manage execution, choose tools, and adapt.
The agent: a discovered path
Ask instead: “Figure out why this order hasn't arrived and resolve it.” Now there's no fixed sequence — each step depends on what the last one discovered.
Same goal, different routes tried one at a time — dead ends in red, the one that works in green.
The developer still sets the tools, permissions and objective — but the agent decides which to use, when, and whether to continue, change course, or stop. Cursor's Agent mode is a clean example: given “add OAuth login,” it searches the repo, edits files, runs commands, and reacts when a test fails instead of following a pre-written edit list.
Names are marketing, not architecture
xAI calls its highly agentic system “Grok Bot.” Tool access alone doesn't make something an agent either — an LLM that always calls the same tools in the same order is still just a workflow. What matters is the decision loop, not the label or the toolset.
And agents aren't rule-free: they still need objectives, permission boundaries, budgets and stopping conditions. The difference is what the rules do:
Bot: rules determine the path. Agent: rules set the boundaries — the agent determines the path.
The real test
Give the system a goal and remove the workflow: can it decide what to do next from what it discovers? Stronger still — can it change its plan when reality disagrees with its assumption? A workflow reports failure. An agent asks what the failure means, investigates, and revises course.
That's the shift from instructions (“do these exact steps”) to objectives (“get this outcome, escalate what needs me”) — and it's what the label “bot” or “agent” should really be measuring.
The boundary begins when the system stops executing the path you designed, and starts determining the path it needs to reach the goal.