Learn how to use generative AI and an AI assistant to create content on RTX Spark
Agents are amazing because you can talk to them in your natural language. While they excel with repetitive tasks, what sets them apart is their ability to run apps you don’t know how to use.
Open your agent and confirm it answers a prompt using your local model. If you don’t have an agent yet, follow the introductory agent playbook first, then come back here.
To do this, you have 2 paths:
Navigate to the Plugins section of your agent (sometimes called Capabilities) and select the apps that you want to connect.
Ask your agent to connect to those apps, and follow the instructions it provides.
One thing you may want to know is that the connection to apps is often done via MCPs - bridges that let agents talk to apps. Because they operate like servers, some apps may require you to turn them on in the app.
Many applications are ready for agents, provided by the software developer or through the developer community.
There’s basically 3 things the agent can do for you:
Adjust something in the app you are working with. For example, you may want to click a button to do something but don’t know where that button is (e.g. Decimate this asset). If you select the asset and then prompt your agent, in many apps it will be able to do it for you.
Automate work that you repeat. For example, you can ask your agent to help turn recurring steps you commonly ask it to do - resizing assets, applying export settings, prepping several files- into a reusable process. You can have it create a skill for instructions and review steps, or a script for a defined operation.
Transform an asset with another app, and bring it back to your app. For example, you may be in DaVinci Resolve and need an adjustment done in your video that you can’t do in this app (e.g. Apply NVIDIA Eye Contact in this video). You can ask your agent to take your current video, process it with NVIDIA Broadcast, and put it back into your timeline.
All steps are done in the same way: simply describing what you want to do to your agent.
Skills are great to teach your agent how you want things done. They are general guidance that the agent should always be aware of when using an app.
Scripts are automated, repeatable workflows - think of a button in a creative app. Once pressed, the exact same process will be done to any asset.
The difference between both is that the Skill will still let the agent think and reason, while the script will create a deterministic pipeline.
You can even combine both ideas! You can create multiple scripts (or tools) for things that you like to automate. And you can create skills that teach the agent when it should use each, and how to chain them.
Let’s learn how to create each:
Describe to the agent what you want done
Once you get the result, give it feedback and ask it to do it again
Once you arrive to your desired result, tell it that this is how you want it, and ask it to create a skill with your desired name.
Create a skill named Product-Concept-Images based on the process we just tested. Ask for the product, colors, materials, lighting, style, intended use, and output folder. Record which ComfyUI workflow to use and how to locate it. Help me refine the prompt, generate an image, and review variations with me before choosing the final version. Show me the skill, then save it using Hermes’s supported process.
I want to create a product concept image. Ask me for the details you need.
Describe to the agent the workflow that you would like done. Make sure that it has clear steps that are repeatable.
Test it and provide feedback.
Once it works, ask the agent to create a script that can be invoked to run this precise workflow, and give it a name.
Create a script named Product-Concept-Variations that accepts a list of prompts and an output folder. Run the ComfyUI text-to-image workflow we just used once per prompt. Save each result with a numbered filename without overwriting existing files, and report failures. Test it with two prompts first.
Run Product-Concept-Variations with these prompts: [list of prompts], and save the results in [output folder]. Ask me for any missing inputs.
NOTE
Agents can get confused if they have a large number of skills and scripts. You can turn off skills that you are not using anymore to improve results. And in agents like Hermes, you can create a bot or profile per workflow with the unique set of skills you need.