Skip to content

Agents

Overview

  • Agents carry out a specific part of the conversation and / or perform tasks like orchestrating between other agents, triggering internal processes and fetching information.
  • Agents carry out tasks through tools provided to them.
  • Agents can be connected to other agents through a mention in the agent's instruction.

Agent Configurations

Description

The description conveys the agent's role in the multi-agent system. Writing a good description is important for other agents to know when to pass control of the conversation to an agent.

Instructions

Agent instructions are the backbone of an agent, defining its behavior. RowBoat Studio's copilot produces a good framework for agent instructions, involving Role, Steps to Follow, Scope and Guidelines. Since agents are powered by LLMs, general best practices while writing prompts apply.

Examples

The agent uses examples as a reference for behavior in different scenarios. While there are no prescribed formats to provide examples in, examples should include what the user might say, what the agent should respond with as well as indications of any tool calls to be made.

Prompts

Prompts attached to an agent will be used by the agent in addition to instructions.

Tools

Tools attached to an agent will be put out as tool calls. The behavior of when to invoke tools can be fine-tuned by specifying corresponding instructions or prompts. Adding examples to agents can also be useful in controlling tool call behavior.

Connected Agents

In the agent instructions, the connected agents are shown with an '@mention'. If the agent mentioned in an instruction (connected agent) does not actually exist, the connected agent's name would show up with an '!' to call to attention.

Model

RowBoat currently supports OpenAI LLMs. Agents can be configured to use GPT-4o or GPT-4o-mini.