Custom agents

Custom agent is an intelligent action that processes natural language inputs to automate complex steps using integrated tool-calling capabilities. It primarily uses integrations as its tool interface, while offering extensibility to use Code as tool, and other native actions like human-in-the-loop task. The agent delivers structured, predictable outputs optimized for seamless integration into downstream automation steps.

Properties

Using Custom Models in Custom Agent (Bring your own bedrock model)

Integrate your desired or custom fine-tuned models hosted in AWS Bedrock with Quick Suite automation workflows.

Before you begin, ensure you have the following:

Step 1: Create a Bedrock Runtime Action integration by following the detailed instructions in AWS service action connectors

Step 2: Set Up Your Automation Group

Create an automation group and connect the integration:

Step 3: Configure a Custom Agent

Add and configure a custom agent to use your fine-tuned model:

Next Steps

Once configured, your custom agent uses the fine-tuned model to process requests according to the instructions you provided. You can now incorporate this agent into your Quick Automate workflows.

Custom agent testing

Custom agent testing enables you to test individual agents independently from the complete automation workflow. This capability helps you validate agent behavior, debug prompts, and iterate more efficiently without executing the entire workflow.

Prerequisites

Start a test

Monitor test execution

During test execution, you can monitor progress in the audit panel on the right side of the screen. The test skips all preceding workflow steps and executes only the selected agent. You get the same logging experience as a full workflow run.

Review test results

After the test completes, review the following information in the Test panel:

Using Custom agent with Build with Assistant

The current tenet for custom agent is it has to be specifically mentioned to consistently get it invoked, here are the things needed in the prompt to make it appear:

- Function names: `use_inline_agent`
- Representation names: "Custom Agent", "Inline Agent" → use `use_inline_agent`
- Generic terms: "agentic skills" → default to `use_inline_agent`

Otherwise, the model is preferred to author the workflow deterministically.

Although, in practice, when no appropriate actions are available, planner might pick custom agent as a workaround. But to consistently invoke custom agent in the workflow, the above phrases are encouraged to use in the prompt.

Examples of agent use cases

Use Case 1: Email Classification and Assignment Agent

Role: You are an Email Categorization and Assignment Agent

Instructions: Follow these steps:

Using knowledge bases with custom agents

In Amazon Quick Automate, you can connect knowledge bases to custom agents to enable AI-powered retrieval and question answering over your organization's documents. By linking a Quick space to your automation group, custom agents can search and retrieve information from the knowledge bases within that space.

Use this for automations that need to reference organizational knowledge — such as answering questions from policy documents, summarizing reports, or classifying content based on reference data.

Knowledge bases index your documents for semantic search, so the custom agent retrieves only the most relevant passages rather than processing entire files. This makes retrieval faster and more accurate, especially across large document sets.

Prerequisites

Link a space to your automation group

Before a custom agent can access knowledge bases, you must link the space that contains those knowledge bases to your automation group. Linking a space grants the automation group permission to access the knowledge bases and files within that space.

To link a space to an automation group:

The space now appears in the automation group's connections list. Custom agents in this automation group can access the knowledge bases and files within the linked space. Other resources in the space are not available to automations.

Add knowledge to a custom agent

After linking a space to your automation group, you can configure a custom agent to use the knowledge bases within that space. The workflow must be in the same automation group where you attached the space.

To add knowledge to a custom agent:

The custom agent can now search and retrieve content from the knowledge bases in the selected spaces when the automation runs.

When you attach a space, all knowledge bases inside that space automatically become available to the agent. You don't need to attach each knowledge base individually. At runtime, the agent queries each knowledge base independently and combines the results in its response.

Writing instructions for knowledge base queries

When a custom agent has knowledge bases attached, it automatically searches and retrieves relevant content based on your instructions. Write instructions that clearly describe what information the agent should find or how it should use the knowledge base content.

Best practices:

Example: Customer inquiry agent with knowledge base

The following example shows how to configure a custom agent that uses a knowledge base to answer customer inquiries based on company documentation.

Setup:

Instructions:

"""You are a customer support agent.

Task: Answer the customer inquiry using information from the knowledge base.

Instructions:
1. Search the knowledge base for information relevant to the customer's question.
2. Provide a clear, concise answer based on the retrieved content.
3. If the knowledge base does not contain relevant information, respond with:
   "I don't have enough information to answer this question. Please escalate to a human agent."

Constraints:
- Only use information found in the knowledge base. Do not make up answers.
- Keep responses under 200 words.
- Include the source document name when referencing specific information."""

Structured output:

{
  "answer": "The response to the customer inquiry",
  "sourceDocument": "Name of the document used",
  "confidence": "high/medium/low",
  "escalationNeeded": false
}