Automating tasks with AI agents
Agentic AI is a new evolution of artificial intelligence that emphasizes autonomous decision-making and task execution with minimal human supervision.
Traditional or generative AI typically requires human prompts or focuses on content creation. However, agentic AI systems, also known as AI agents, are designed to do the following tasks:
- Independently make decisions
- Take actions
- Solve complex problems
- Interact with external environments
These systems are proactive rather than reactive, and can pursue goals, plan multi-step tasks, and adapt to real-time changes without constant human input.
Capabilities
You can build and deploy AI agents to automate tasks and make your applications more flexible and dynamic. Agentic applications allow a foundation model to function as an agent that controls the flow of interaction with the user.
In a real-world workflow, the foundation model within the agent picks one or several external or custom tools based on prompts submitted by the end user. The model then uses a process called tool calling, which is also referred to as function calling, to use one or several tools to process the input and generate a response.
For example, a restaurant reservation app might need to collect the following information, and then create a reservation:
- Location
- Number of guests
- Date and time
The order in which the information is collected might matter only in some locations, where the number of guests allowed per table is limited, or where more seating is available outside in fine weather. You can account for these conditions in the code, and let the agent determine when to call extra tools, such as a third-party weather app, before the agent completes a reservation request.
You can configure the agent to make decisions and perform tasks on behalf of an end user. You define the parameters of the interaction between the agent and the end user. The agent uses the parameters to decide the next best step based on the current state of the interaction.
When you build agentic applications, you must configure the following core components:
- Foundation model
-
When you build an agentic workflow, choose a foundation model that meets the following requirements:
- Handles chat tasks
- Supports tool calling
- Responds well to chain-of-thought prompting techniques and chooses the next action
- Has a good performance on reasoning benchmarks
To find the list of foundation models that you can use in the Agent Lab, see Configuration options in Agent Lab. To programmatically list the foundation models that support tool calling in watsonx.ai, see Agent-driven chat.
- Framework
-
Select an AI agent framework you want to use to create, deploy and, manage your agent. You can only use the LangGraph framework to build agents in the Agent Lab. You can use various alternative frameworks if you choose to build an agent locally, and then deploy the agent in watsonx.ai.
- Architecture
-
Select the architecture that implements agentic AI reasoning. You can build agents with the ReAct (Reason + Act) technique only. ReAct creates a continuous loop of reasoning and action in which the outcome of each step informs the next.
- Instructions
-
Specify specific instructions for your agent that are used in the system prompt for the foundation model. The instructions can include using a specific language, date or time format, user greeting, or an external tool as an information source instead of a foundation model's knowledge base. For details about how to craft effective prompts, see Prompt tips.
- Tools
-
Specify external tools that can retreive additional information to use to ground prompts submitted to the foundation model. Choose from the following types of tools:
- Provided tools that are ready to use in watsonx.ai.
- Custom tools that you create for processing specialized input formats and that generate a specialized result. See the following examples that demonstrate how to create and use custom tools:
- From the UI: Creating a custom tool in Agent Lab
- From the API: Example tool-calling API request.
Ways to work
You can use the following graphicals tools to build agents in watsonx.ai:
- Agent Lab: A graphical tool from the watsonx.ai UI to build one agent at a time.
You can use the following programmatic methods:
- Use the watsonx.ai chat API with foundation models that support tool calling to build agent-driven chat workflows. For details, see Building agent-driven chat.
- Develop and test an agent locally before you deploy the agent in watsonx.ai. For details, see Agents in the Developer Hub documentation.
Samples
The following table contains samples that demonstrate how to build agents that can be deployed in watsonx.ai:
| Example | Format | Description | Link |
|---|---|---|---|
| Complete agent built in watsonx.ai | Agent | Sample agent that generates recipe ideas based on a list of available ingredients by using the foundation model's knowledge base and external tools. | Sous Chef sample agent |
| Agents that use various frameworks | Python | Sample agent templates you can use to develop and test your agent locally before deploying the agent. The templates include the necessary code to setup the agent to be deployed as an AI service in watsonx.ai. For details, see Coding and deploying AI services with templates. | Supported agent templates |
Learn more
Parent topic: Gen AI solutions