Skip to content

agent.md

Latest commit

 

History

History
74 lines (55 loc) · 3.01 KB

agent.md

File metadata and controls

74 lines (55 loc) · 3.01 KB
Raw

Agent Management

Use Shell Scripts

We have provided some helper shells for your convenience.

When you use these scripts, make sure you have started the api server in localhost.

For example, we want to create an agent, id is example, first cd into the directory:

cd scripts

Create Agent:

sh create.sh example

Now you have a blank agent, let's add features to it.

Export Agent:

sh export.sh example

Edit the agent config file: example.yaml. Then import it.

Import Agent:

sh import.sh example

Test Your Agent:

You can test your agent using API calls or the IntentKit Sandbox UI. The UI provides a visual chat interface for interacting with your agents.

You can use the hosted version at https://intentkit.crestal.dev/ without needing any local setup, or you can set up your own instance by following the instructions in the IntentKit Sandbox UI repository.

Advanced Agent API

You can visit the API Docs to learn more.

The Prompt

In agent config, there are 5 fields about prompt, they are purpose, personality, principles, prompt, prompt_append.
IntentKit will compose purpose, personality, principles, prompt to system_prompt.

LLM Interaction

The models cannot remember anything, so every time we interact with it, we have to provide all the background information and interaction context (that is, additional knowledge and memory).
What we send to the large model looks something like this:

  • System: system_prompt
  • User: conversation history
  • Assistant: conversation history
  • ...
  • User: conversation history
  • Assistant: conversation history
  • User: currently being said
  • System: prompt_append (Optional)

The content of the system role is to inform the AI that it is being addressed by an administrator, so it should not treat you like an user. However, your permissions are not necessarily always higher than those of regular users; you simply have the advantage of being the first to set various rules for the AI to follow according to your logic.
For example, you can tell it that the system role has the highest authority, and if the user role requests an action that violates the rules set by the system role, you should deny it.

Prompt and Append Prompt

Writing the initial prompt is a broad topic with many aspects to consider, and you can write it in whatever way you prefer. Here, I will only address the prompt_append.
We've found that if you emphasize the most important rules again at the end, it significantly increases the likelihood of the AI following your rules. You can declare or repeat your core rules in this section.
One last tip: the AI will perceive this information as having been inserted just before it responds to the user, so avoid saying anything like "later" in this instruction, as that "later" will never happen for the AI.