Configuring Now Assist AI agents
Summarize
Summary of Configuring Now Assist AI agents
Now Assist AI agents enable ServiceNow customers to automate workflows by configuring AI agents that execute tasks using mapped tools and business logic. These agents leverage record context and searchable content combined with large language models (LLMs) to plan and suggest next best actions toward specific goals and outcomes. Ensuring that your data and knowledge base are accurate is essential for optimal agent performance.
Show less
Prerequisites for Configuration
- Define the types of tasks your AI agent workflows should handle.
- Understand the general flow of your agentic workflows and agents.
- Use agentic tools with clear, well-written descriptions.
Key Configuration Elements
- Base Plan: Sets initial instructions for the AI Agent Orchestrator at the workflow level.
- Role: Establishes the AI agent’s identity, including reasoning prompts and an auto-generated proficiency description that details capabilities based on assigned roles and tools.
- Instructions: Provide clear, step-by-step operational directives for the AI agent.
Configuring Tools for Agentic Workflows
Tools should be configured with a single, well-defined purpose to improve agent reasoning and execution efficiency. Multipurpose or multi-mode tools can confuse agents and degrade performance. Tool descriptions must clearly outline:
- Tool functionality and intended use.
- Specific scenarios and workflows where the tool applies.
- Situations where the tool is not applicable to avoid confusion.
- Relevant terminology to clarify context.
Error messages generated during tool execution help AI agents learn and adjust, maintaining workflow accuracy.
Dynamic Orchestrator
When more than 8 to 10 AI agents exist for a workflow, the Dynamic Orchestrator selectively identifies the appropriate agents during planning to optimize performance and accuracy.
Invoking AI Agent Conversations
Use the AI Agent Background Channel and associated Provider to initiate AI agent executions from the Workspace. Conversations are triggered via an API call and recorded in the Execution Plans table, where you can monitor execution tasks, messages, and tool usage. AI Agent Studio Testing allows detailed review of execution steps and decision logs.
Interactive vs. Non-interactive AI Agents
- Interactive agents prompt users for information during execution fallbacks and re-trigger workflows as needed.
- Non-interactive agents do not request user input during fallbacks; instead, they adjust prompts dynamically and display results or error messages in the Now Assist panel or Virtual Agent.
The execution mode is configurable via the Execution Plans table, allowing you to select the appropriate interaction style for your use case.
Configure the Now Assist AI agents to execute agentic workflows with AI agents and mapped tools.
Configuring AI agents
- Prerequisites
- By making a plan, you can improve your AI agent performance and result quality. When you have a solid foundation of what you want to build, you can minimize creating redundant agents and maximizing the efficiency
of your existing ones. Before you send instructions to your AI agents, make sure that you follow these prerequisites:
- Have a good idea of the different kinds of tasks that your agentic workflow should be able to handle.
- Understand the general flow for your agentic workflow and agents.
- Use agentic tools with well-written descriptions.
- Configurable elements
- Instruct the agentic workflows and AI agents through the following elements within the framework:
- Base plan: Instructions to the AI Agent Orchestrator for the initial planning procedure that is configured at the agentic workflow level.
- Role: Clear identity of the AI agent that includes these elements:
- Agent reasoning: When a role is added to each reasoning prompt, it provides a sense of identity to the content that is generated by the LLM.
- Agent proficiency: An LLM-generated description of what an agent is capable of, including the content from the role, instructions, and the descriptions from the tools that
are assigned to the AI agent.Note:The agent proficiency is auto-generated.
- Instructions: Clear directives for the AI agent. Write instructions as a step-by-step algorithm that describes the operational flow for the AI agent.
Configuring the tools for the agentic workflows
- Functionality
- What an AI agent contributes to the agentic workflow. Configure the tools with a single purpose. Multipurpose tools can cause a problem for the agents for the following reasons:
- Multipurpose tools are harder for the AI agent to reason through and determine when to use the tool. If a tool can be used for more than one purpose, the AI Agent Orchestrator has to determine which purpose is most applicable, which can decrease your AI agent's performance by increasing the runtime.
- The tool description must be comprehensive enough to account for all the scenarios for the usage of the tool that is being defined.
Note:Don't use tools that can operate in different modes. Instead, configure your tools as the solution to a singular problem for a scenario. - Tool description
- Natural language descriptions that describe the utility provided by the tool. Make sure that you define the scope and limits of the tools clearly to help ensure that the tools are picked for the appropriate
scenarios in the following ways:
- Provide a description of what the tool is supposed to do.
- Describe the scenarios where the tool can be called. Include the specific agentic workflows and tasks where the tool and its functionality can be used.
- Explore the scenarios where the tool is explicitly not useful but an AI agent can confuse the tool as being useful.
- Explain the terms that are being used in the preceding cases. For example, if you have a tool for assigning a role to a user, you must explain what the role is in the agentic system of the given instance.
- Error messages
- An AI agent operates through trial and error. For example, an error message about an execution that accidentally ran incorrect tools can help the AI agent reach more valid conclusions in the future. Error
messages offer an AI agent a chance to reflect and explore other options.
Understanding the scenarios where the tool can go wrong can help the AI agent with keeping the execution on track.
Dynamic Orchestrator for agentic workflows
When there are more than the ideal 8 to 10 AI agents available for an agentic workflow, then the dynamic orchestrator helps identify the agents that need to be considered for executing that agentic workflow in the planning phase. The dynamic orchestrator helps map the right agents to the user's agentic workflow to provide better performance and accuracy.
Invoke Conversations with AI Agent Background Channel
The AI Agent Background Channel helps you to invoke AI Agent or agentic workflow execution from the Workspace. Use the AI Agent Background Channel associated with the AI Agent Background Provider to invoke conversations. The AI Agent Background Provider is based on the Custom Adapter Framework from Virtual Agent. For more information, see Configure a provider for your custom chat integration.
Create a channel identifier in the Provider Channel Identities table [sys_cs_provider_application] to add any additional conversational capabilities to your own provider application and get a new inbound ID that allows for customization. For more information, see Create a channel identifier for your custom chat integration.
To start a conversation, trigger the flow using the sn_aia.AiAgentRunttimeUtil().startAiAgentConversation(request) API in the Script Include (sys_script_include) of the AIAgentBackgroundProvider and select Run Script. When the Script execution status indicates Success, the conversation begins in the order of utterances defined in the Script.
Conversations that are invoked for executing an AI agent are logged in the Execution Plans [sn_aia_execution_plan] table. Open the conversation record to confirm the device type as AI Agent Background. Open the execution record to see the Execution Tasks, Messages, and the Tools Executions used to execute the AI agent.
You can also see the entire execution steps on the AI Agent Studio Testing page by copying the execution plan record's [sys_Id] and testing it. On the Chat responses tab, in the AI agent decision logs, you can see the AI agent details and the tools it used to resolve the issue.
Interactive and Non-interactive AI agents
The Interactive AI agents reach out to users for information when there is a fallback in the execution process, and the AI agent re-triggers the flow.
The Non-interactive AI agents do not reach out to the user at any fallback stage in the execution process. When the AI agent needs user information, it takes the dynamic prompt approach using the ReAct layer, where the prompt of the ReAct will change based on the execution mode of the AI agent or agentic workflow. Therefore, in the Non-interactive execution, the reach fallback options do not have to collect input from a user as a fallback option. However, the output of the AI agent or agentic workflow will still need to be presented to the user, and in any execution failure scenario, a message in the Now Assist panel or Virtual Agent is shown.
To implement the Non-interactive execution, the Execution Mode field is added in the Execution Plans [sn_aia_execution_plan] table, where the execution mode can be Interactive or Non Interactive based on the given runtime parameter.