Chatbots and agents

A chatbot receives a message and returns a response. An agent can inspect external state, choose actions, use tools, read the results, and continue toward a goal.

The boundary is not perfectly sharp. A chatbot may search the web once. An agent may stop after one tool call. What matters is how much of the sequence the system can choose and execute without you naming every step.

Compare these requests:

Explain why this date test is failing.

Inspect the failing date test, fix the bug, and verify the change.

The first can end with an explanation. The second needs to read files, run a command, edit code, and run the test again. It also needs permission to change the repository.

The model does not own the tools

The model proposes a tool call, such as read this file or run this test. The host application checks the request, applies permissions, executes the tool, and returns the result.

This distinction matters. The model does not magically possess your filesystem or account permissions. The surrounding application decides which capabilities exist and which calls require approval.

A small agent trace

Goal: fix the failing date test
Observe: read the failure
Act: inspect the date formatter
Observe: find a timezone assumption
Act: edit one function
Check: rerun the focused test
Stop: the test passes and the diff is in scope

Each observation changes the context for the next decision. The agent can recover when the evidence contradicts its first idea.

More agency means a larger blast radius

Explaining how to send an email changes no external state. Drafting one creates text you can review. Sending it reaches another person and may not be reversible.

As agency grows, pay attention to available tools, exact permissions, approval points, cost limits, stopping conditions, and verification. Autonomy does not remove responsibility. It increases the importance of boundaries.

Try this

Take one task you give a chatbot. Rewrite it as an agent task. List every extra tool, permission, side effect, and check the agent would need.