Skip to main content

Retool Agents overview

Learn about Retool Agents.

Version: Cloud and 3.234 Edge

Retool Agents is a product that makes it simple for builders to automate work using large-language models (LLMs) by creating agents. An agent is a system that can complete or delegate tasks based on LLM reasoning. You can invoke agents using chat and email, evaluate agent performance, and even call them directly from apps and workflows.

Agent process​

Agents call tools—for example, workflows, functions, or other agents—to gather information and complete or delegate actions. When invoked, an agent:

  1. Receives a task, or input, as natural language. Tasks are provided as input to agents via written instructions.
  2. Uses an LLM to decide whether to respond to the input, get more information, or take action in another system.
  3. Provides the result of the tool call back to the LLM, and the LLM reasons in an open-ended cycle called an agentic loop without a pre-defined stopping point.

Agentic loop steps can occur up to the maximum number of iterations specified in Retool Agents. A run is a single agentic loop defined by the agent, its input, and its output.

An agent acts according to the instructions it is provided. The LLM can only make two decisions: to continue calling tools, or to provide a final answer. It can call an LLM, execute tool calls, and provide the data from the tool calls back to the LLM in a loop until the LLM decides the task provided is complete.

Invoke an agent​

You can invoke an agent using any of the following trigger methods:

An agent only runs when invoked, such as with the Agent Chat component, and cannot be directly configured to run automatically. A workflow, however, can run automatically using a schedule or in response to webhook events. You can combine a workflow and the Invoke Agent block to automatically invoke an agent based on your requirements.

For example, you can create an agent using the Meeting Prep template that sends you a pre-meeting email with event and attendee information. On its own, the agent cannot be automatically invoked. To automate this, you would create a workflow that includes the Invoke Agent block and is configured to run periodically. When an upcoming event is found, the workflow invokes the agent. The agent, in turn, then sends the necessary information to prepare for the meeting.

Loading diagram...

Agent architecture​

The architecture of an agent consists of:

  • Orchestration layer: Manages memory, state, reasoning, and planning.
  • Model: The LLM that the agent utilizes.
  • Tools: Bundled capabilities that allow an agent to perform an action (e.g., access an API, a website, etc).

Agents use frameworks like ReAct (Reasoning and Acting) to alternate between thinking and taking action.

Loading diagram...

When to use agents​

Agents work well for open-ended, dynamic tasks, but they are not optimal for well-structured tasks that can be encoded into discrete steps. For well-structured tasks, it may be more beneficial to use agentic workflows. The following diagram is a visual representation of the different types of GenAI and Agentic workflows that can be built using Retool Workflows, vs. the flow of a Retool Agent.

Loading diagram...

For more information about the differences between agents and agentic workflows, refer to the Agentic Workflows conceptual guide.

Why Retool Agents​

Retool Agents gives builders a central place to create, test, deploy, and monitor AI agents.

Some of the key features that Retool Agents provides are:

  • Intuitive agent-building experience:
    • Form-based UI for agent configuration.
    • Built-in chat experience.
    • Integration with existing Retool components.
  • Thoughtful tool functionality:
    • A set of Core tools for out-of-the-box functionality to get your agents up and running faster (Retool Email, Retool Storage, Web Search, Code Execution, etc.).
    • Ability to create custom tools.
    • Natural language tool generation with the Config Assistant and Function Generator.
    • Ability to connect to an MCP server to pull in third-party tools.
  • An evaluation framework:
    • Ability to test against datasets.
    • Multiple included reviewer types (exact match, LLM as judge, etc.).
    • Ability to score agent runs with an eval.
    • Ability to compare two runs side-by-side.
  • Monitoring and observability:
    • Visibility into agent runs, token usage, and cost.
    • Visualization of agent behavior.
  • A Human-in-the-Loop (HITL) component:
    • Consent mechanisms for tool execution.
    • Chat to test.
    • Version history and change tracking.

Additional resources​

To continue learning about agents, refer to the following resources for more information.