Agent Routines

Owned by Alex Roytman
Last updated: Dec 17, 2023
3 min read

Agent Routines in Profound AI are designed to enable AI agents to execute a variety of automated tasks and actions, ranging from simple data retrievals to more complex operational procedures such as processing orders. This feature allows for seamless interaction between AI agents and your business systems, enhancing efficiency and user experience.

Agents decide on their own when a particular routine is pertinent in the context of a user interaction. Therefore, it is important that the appropriate routine name, description, and its inputs are provided, as well as sufficiently described using natural language. To use Agent Routines, the large language model you choose must have been trained or fine-tuned for function calling.

Managing Agent Routines

To create, edit, or remove agent routines, simply switch to the Routines tab in the IDE when working with an agent, and then use the appropriate toolbar icons.

In the General Info tab, make sure to provide a clear name and description to describe the routine’s capabilities. The AI Agent reads this information to decide when to call the routine.

Mark a routine private if it is intended to be used by other routines or for invocation from from Agent Instructions. When a routine is marked private, the AI model itself does not receive any information about it and is unable to initiate a call to that routine directly.

Adding Input and Output

You can define Input parameters and the routine Output in the bottom section of the IDE.

The parameters can be repeated as arrays and nested to as many levels as necessary. For example, a routine for placing an order may require customer information as well as a list of order detail lines with nested information, such as product id, product configuration, and quantity for each detail line.

It is important that Input parameters are defined to that level of granular detail, with Examples and Descriptions provided to guide the LLM in generating the appropriate parameters.

Defining Output may not require the same level of granularity since the LLM can potentially infer the meaning of the data from the output response itself. For example, if a list of records is sent as output, it may not be necessary to define each record column when the data itself contains the column names as part of its JSON response.

Adding Routine Logic

Routine Logic can be specified as either Node.js code or as a set of Low-code steps.

Low-code steps are configured using a drag-and-drop interface in the center of the IDE. They allow you to create functionality just by answering questions. The steps are defined from a set of plugins, which facilitate various functions. For example:

  • Access to any configured databases for retrieval, updates, summarization, and more

  • Web Service Consumption

  • Calling ILE programs and service programs

  • Calling Operating System Commands

  • Custom SQL or Node.js logic

  • Conditional and repeating blocks

  • And more

More details about using Low-code can be found here.