Skip to main content

​
Multi Skills / Dynamic Skills

Multi Skills or Dynamic Skills are a way to give your agent the ability to perform multiple flows, each with its own context + tools.

​
How it works

First you define a base prompt which defines the agent persona:
  • Who the agent is
  • Which company is it working for
  • Identity + language
  • In scope + out of scope rules
Then you define the skills. Each skill consists of:
  • A prompt which defines what to do when using said skill
  • A list of tools to help accomplish said task
The agent will call the switch_skill(skill_name) tool to switch to the desired skill when necessary.

​
Under the hood

Under the hood, each time the agent calls switch_skill(skill_name), we update the underlying system prompt of ‘gpt-realtime’ to the prompt of the desired skill + the tools of the desired skill. The agent does remember the conversation history it had before switching the skill.

​
Why is it better?

Due to the fact that each prompt is focused on a single task, the prompt is shorter and you can fit more examples in the prompt. This leads to much better instruction following.

​
Naming tools in skills

Try to namespace your tools. For example, if you have a skill called “technician” which is supposed to help with booking a technician, you should name your tools accordingly:
  • technician_book()
  • technician_cancel()
  • technician_rebook()

​
When not to use multi skills

For example if you have a long QnA document, it is better to write an answer(question) tool that will answer the question based on the QnA document, than making a skill containing the QnA document.

​
Base tools

Base tools are tools that are always available to the agent no matter which skill it is in.