Claude Agents

 

In Claude code you can agents.  Now what is an agent?   You can think of it as a smart assistant.  You can create an agent, give it a goal, then have it go at it working independently then return the results. 

In this article I am going to keep it very basic and simple and focus on just creating your first agent.

 

See https://code.claude.com/docs/en/sub-agents#quickstart-create-your-first-subagent [1] for more detailed information

Making your first agent

Open up Claude code and run the command /agents

> /agents

 

You will see this tabbed menu with the tabs running and Library.

You can see I have no agents running right now.

 

If I go to Library and go down you see I have an option to create a new agent.

Click enter

Choose a location.
In this case I am going to choose my personal directory vs the project.

 

Now choose Generate with Claude

 

Now we start our conversation to describe what this agent should do

Here is an example of what you might normally do with an agent.

A code improvement agent that scans files and suggests improvements for readability, performance, and best practices. It should explain each issue, show the current code, and provide an improved version.

 

(taken from https://code.claude.com/docs/en/sub-agents#quickstart-create-your-first-subagent [1])

But we are not going to do normal we are going to do simple so you know how to set up and use the tool.

Let’s create a price checker.  I am in the market for a Western Digital Pro 20TB hard drive (model number WD202KFGX).

My prompt

A price checker agent that will check the current price for a Western Digital Red 20TB hard drive model number WD202KFGX

 

It will start to generate the agent

 

Now you need to select tools…. Why?

Well normally Claude is very cautious.  When you ask it to do something it asks for your direct permissions to do that action (in many cases)
For an agent we would like them to be autonomous (work on their own with no…  or very little interaction from us)

So, what can we do here? 

We could select All Tools then it can use all the tools.  This would make it very powerful but also, potentially, very dangerous.  In my case I need it to be able to read from web sites… So what permissions should I give it.

Let’s select
Read-only tools

 

When I press enter I see it change to 20 of 23 tools selected.

Now go up and select Continue

 

Select the model.  I will select Opus

 

Choose a color.  I will choose Red

 

Now for memory.  I am going to select User Scope.

But why?  And what is this?

Well basically… The first time this agent kicks off ever it will have to keep track of a few things as it tries to accomplish its goal.  Once the goal is accomplished, if you have no persistent memory then the next time you run the agent it will be like running it totally fresh without any memory of the last run or runs.

 

However if we allow it to save a memory it will have that next time it runs (it can add/tweak it on every run as well)

So for now lets set it to User Scope

 

There are our results.

Press Enter to Save

 

Now we can see that this agent exists.

 

And if I look in my file system

> ls ~/.claude/agents

 

There is an agent Mark Down file there.

Also if I look here

> ls ~/.claude/agent-memory/

 

I can see there is a memory folder for this agent but there is nothing there yet.

 

 

 

Running the Agent

Now that the agent exists how can I run it?

Use the @ followed by the agent name (You can use tab completion)
Or you can scroll down through all the options

 

> @wd-red

(then press tab)

Now run it.

It came back and asked for permissions.

 

And another ask, and another, and another.  I think I have some tweaks to make.

 

OK I got a result..  (A bit off... but a result)

But I had to interact with it let me see if I can fix this.

> For the agent @/home/patman/.claude/agents/wd-red-20tb-price-checker.md  I need to make some permission adjustments.   When I last ran it I had to give it permission to access web sites and also to adjust its own memory.  Please adjust   the agent so it has the permissions to do this and does not need to ask for it next time.

 

You can see it update the settings.json file.

Let me do one more tweak.

> Update the @/home/patman/.claude/agents/wd-red-20tb-price-checker.md agent so that when it finds the best price it opens the URL using the open command line command

Now let’s try it again

> @wd-red

(then press tab)

 

OK.. failure

 

A few more back and forth conversations to get it working….

 

OK it worked without any intervention on its own.  It still needs some tweaks but it works.

So why use an agent vs just using a Skill?    They both have Mark Down files (instructions).  So why have agents?

 

1.      Context -
     Agents have their own context so they can focus on their mission

2.      Memory -
     Agents can keep their own memory between sessions

3.      Parallel -
     Agents run in parallel.  So you can continue your session while they run in the background (They could be running for hours to solve a problem)

 

 

I have not gotten into having multiple agents running or even agents coordinating (this can be very powerful).

But for now I just wanted a simple write up on getting something working.

 

 

References

 

[1]       Quickstart: create your first subagent
            https://code.claude.com/docs/en/sub-agents#quickstart-create-your-first-subagent
            Accessed 06/2026