MAX’S NOTES
AI chat, skills or agents: which does your task need?
These terms describe different parts of the work. A fictional family trip makes the distinctions easier to see, without promises of complete autonomy.
Chat is a way to interact with a system. A skill describes a reusable way of working. An agent can choose steps and use available tools to pursue a goal. They are not three mutually exclusive purchases: an agent may use a skill within a chat.
“Agent” has several meanings. Anthropic distinguishes predefined workflows from systems that dynamically direct their steps and tool use. [1] I use that distinction as a useful guide, not the only possible classification.
Chat: work through a specific question
Imagine a fictional family weekend trip.
In a chat, you can explain your preferences, ask for options and refine the plan. You guide the conversation by adding a budget, changing the departure time or ruling out a route.
A conversation alone does not imply current price searches or booking capabilities. Check which tools the system has and what it is allowed to do.
Skill: keep a process you want to reuse
Suppose you want every trip assessed in the same way: travel time, pace, family interests, spare time and unknown conditions.
That process can become a reusable instruction. In the open Agent Skills format, a skill is a folder containing SKILL.md and optional references, templates or scripts. The environment must support that format. [2]
A skill explains how to work. But “check prices” does not provide access to websites, and “every Friday” does not create a functioning scheduler.
Agent: coordinate steps using tools
For our fictional trip, an agent might find accessible sources, compare options and assemble a plan. It needs the relevant tools and permissions.
Set the boundaries first:
- Which sites and information may it use?
- May it only compare, or also place an order?
- Where must it stop and show a result?
- Where will the work be saved?
I would start with preparing options without payment or booking. That lets you judge the result before deciding which actions to delegate.
Choose according to the work
For a one-off question, start with a conversation. When a task repeats and you know what a good result looks like, capture the process as a skill. When it requires several steps and tool actions, consider an agentic process.
All three still need checking. None of the names proves that a system runs continuously. An “agent” on a cover is not evidence that someone has actually started a shift.
Three questions before starting
Ask what the system receives, what it can actually do and what evidence of completion it will show. Those answers matter more than an impressive label.