- Getting started
- Prerequisites
- Building agents in Studio Web
- Building agents in Agent Builder
Agents user guide
Tools
Tools offer agents both access to critical context from data stored in business applications, and the ability to execute actions based on objectives outlined in the prompt. Tools are how the agent's reasoning and planning can turn into action.
Without giving an agent explicit access to a tool, the agent is not able to execute functions (outside of Context Grounding and Escalations, which are separate components of an agent). When adding a tool for your agent, provide a focused description of its purpose and capabilities, along with an explanation of when to use it in the system/user prompt. This ensures the agent understands the tool's functionality and can use it effectively in appropriate scenarios.
Agents use Integration Service-based activities from connectors such as Box, Coupa, Microsoft Teams, Salesforce, Slack, and Zoom. Using secure, governed Integration Service connections and the field configurations described in the following sections, an agent is able to make API calls directly to a target business application. Note that not all Integration Service activities are available in Agents.
Configuring activities to act as tools
- From the Design tab, under the Tools section, select Add tool.
- Search for activities using the search bar or select a connector from the list to see what activities are available.
- Once you add an activity as a
tool, you can:
- Update its description.
- Select a preexisting Integration Service connection or create a new one.
- Configure the field
input options to help your agent learn how to use the activity.
Integration Service connections can be added and managed as in Studio. These secure, governable connections are monitored by Integration Service, which notifies users of any breaks or invalidations.
Each activity comes with a default description, which you can customize to better guide your agent during runtime. While optional, this step enhances output quality and helps the agent understand how to use the tool for specific use cases.
Agents inherently know the information required for successful activity execution, eliminating the need for explicit input or output specifications. By default, agents infer necessary field values. However, you can modify the input method to Static Value or Argument for using specific data or retrieving data from agent input arguments.
Defining activity input methods
For activity fields, you can select one of the following input methods:
- Agent (default) - When selecting this option, the agent infers the field value at runtime using the Description field and other agent reasoning pertinent in determining to use the tool.
- Argument - When selecting this option, you set the field to an agent input argument. You must first define the argument. Use arguments for passing in runtime values from a workflow that precedes the Run Job activity.
- Static Value - When selecting this option, you set the field to a specific, static value every time the agent runs.
Depending on the activity, certain fields may require static values. For example, a specific Slack user you want to set in the Send as field. Other fields default to being inferred by the agent.
Retrieving context from the web
Web Search and Web Reader are two first-party tools you can add to your agents to supply up-to-date context from the web or a public domain. You can use the activities for questions regarding real-time data, such as stock prices, the time, weather forecasts, or data available on a specific URL at any give time.
The Web Search tool executes a web search using natural language. Like other tools, the Search field can be inferred from instructions in the system and user prompt. For instance, you can tell your agent to get details on industry trends and competitor activities which would be used to construct a forecasting report. By offering the instruction in your prompt and adding the Web Search tool, the agent knows it needs to use the tool to access this critical, up-to-date context.
Web Search executes a full summary and provides it to the agent based off of the top hits. The output includes the URLs where the information was sourced from.
The Web Reader tool gives you the ability to extract text from a specific URL and add that as context for your agent. For instance, you have a specific parts supplier and you want to go to their web page and extract current prices for those parts. By supplying the URL directly or through the prompt, you ensure that only information from that individual URL is used for surfacing the supplier's prices for specific parts (or any other information contained in that URL).
By combining both Web Search and Web Reader, you give your agent the ability to execute research in the public domain and then use the cited sources and URLs to extract specific details.
An agent can run select processes published to your tenant. This is done on a per-process basis.
An agent can run selected processes published to your tenant. This is done on a per-process basis.
- From the Design tab, under the Tools section, select Add tool.
- Select the Process tab
and search for processes using the search bar or select one from the
available list.
To access a process as a tool, you must first publish and deploy it to a shared Orchestrator folder you have access to. The process must also conform to the supported input and output types.
Once you add a process as a tool, use the system prompt to describe its function and when it should be used. For any arguments, you can provide a description to help the agent use it effectively.
The following input and output argument types are supported for processes. If these limitations are not observed, processes may be inaccessible or fail.
"System.String", "System.Char" "System.Byte", "System.SByte", "System.UInt16", "System.Int16", "System.UInt32", "System.Int32", "System.UInt64", "System.Int64" "System.Double", "System.Single", "System.Decimal", "System.Boolean"
An agent can run other agents published to your tenant.
- From the Definition panel, under the Tools section, select Add tool.
- Select the Agents tab and search for agents using the search bar or
select one from the available list.
To access an agent as a tool, you must first publish and deploy it as a process to a shared Orchestrator folder.
- Once you add an agent as a tool, use the system prompt to describe its function and when it should be used.
