Introduction
In my previous articles, we explored how AI Agents can transform business operations from a bakery CEO's perspective and showcased data workflow. Let's dive deep into the technical implementation using AWS Bedrock AI Agents defined via CloudFormation and SAM. This guide will focus on the key components of the AI Agent configuration and deployment process, including SAM AWS Lambda to grab data, AI Agent tooling, and AI Agent itself.
This article will give you a working template for building multiple custom AI Agents and showcase how to call them programmatically from your custom user interface.
Note 1. If you dont feel comfortable with the code, I still recommend you check the AI Agent definition to understand the moving parts. It is human readable and I hope, logical: https://github.com/javatask/ai-agent-ceo-fin-advisor/blob/main/bedrock-agent-stack.yaml
Scroll down to skip permission policy section.
Note 2. Source code is avaliable here: https://github.com/javatask/ai-agent-ceo-fin-advisor/tree/main
Prerequisites
Before starting, ensure you have:
AWS Account with Bedrock model access enabled
AWS CLI configured with appropriate permissions
AWS SAM CLI installed
Basic understanding of CloudFormation and Lambda
JQ was installed for JSON processing, and I used it to grab AWS Lambda arn from SAM deployment
Implementation Overview
Our implementation consists of two main CloudFormation stacks:
Finance Tool Lambda (The Intern's Laptop):
Bedrock AI Agent Configuration (The Intern's Guide):
It is best practice to manage the life cycle of your Interns via the OpenGitOps approach, so any degradation in performance or need to be deployed in other regions can be easily made via CI/CD. Ad-hoc changing tooling, instructions and/or interns are OK for development, but for production release, put the Intern version under git control and deploy via CI/CD pipeline.
Understanding AI Agent Components
1. Intern Laptop or Action Groups
The AI Agent needs specific tools to perform tasks. In our case, I defined two functions only analyze_industry_performance and send_email. The goal of analyze_industry_performance the function is to take three parameters. industry, date_from and date_to - these parameters are taken from chat input. For example, in my previous article, I gave the Intern a question How did we do with the hospitality business in the first half of the year?. Our Intern was capable of mapping Hospitality to function parameters industry=hotels and setting valid date_from and date_to. After calling analyze_industry_performance function, the Intern can make HTML or Markdown formatting to send a report via email or show it as a Markdown output.
It is important to give a detailed description of the function and its parameters so the Intern is capable of mapping CEO intent to function or even asking additional questions that are required to build a report based on CEO intent.
ActionGroups:
- ActionGroupName: financial-analysis-actions
Description: Financial analysis and reporting functions
ActionGroupState: ENABLED
ActionGroupExecutor:
Lambda: !Ref LambdaFunctionArn
FunctionSchema:
Functions:
- Name: analyze_industry_performance
Description: |
Analyzes financial performance metrics across a network of bakeries, providing insights into bookings, billings, and industry-specific trends.
Industry Parameter Values Valid values:
'schools' - Educational institutions and campus dining
'cafes' - Coffee shops and small eateries
'shops' - Retail bakery outlets
'factories' - Industrial/manufacturing facilities
'restaurants' - Full-service restaurants
'hotels' - Hospitality sector
Date Range Parameter Format: ('YYYY-MM-DD', 'YYYY-MM-DD')
Valid ranges:
Full year: date_from:'2024-01-01', date_to:'2024-12-31'
Q1: date_from:'2024-01-01', date_to:'2024-03-31'
Q2: date_from:'2024-04-01', date_to:'2024-06-30'
Custom: Any date range within 2024
Parameters:
date_from:
Type: string
Description: date from in format YYYY-MM-DD
Required: true
date_to:
Type: string
Description: date to in format YYYY-MM-DD
Required: true
industry:
Type: string
Description: industry for each need to do analysis
Required: true
- Name: send_email
Description: Function to send an email to CEO with arbitrary payload
Parameters:
html_body:
Type: string
Description: HTML text of the email
Required: true
subject:
Type: string
Description: subject of email
Required: true
2. Your Guide to Intern or Agent Instructions
Instructions are crucial for AI Agent behaviour. The main point here is to guide the Intern on how to use a laptop and data access.
Instruction: |
You are an experienced financial analyst specializing in business finance.
Your task is to prepare ad-hoc reports for the CEO based on the financial data
provided and the specific request made. Follow these instructions carefully
to produce a comprehensive and insightful report.
First, you will be presented with the financial data by using the function analyze_industry_performance
Next, you need to work on CEO's specific request.
If required, use the send_email function to send the HTML version of the report to the CEO's email.
Analyze the financial data in the context of the CEO's request. Consider the following steps:
1. Identify the key financial metrics relevant to the CEO's request.
2. Perform necessary calculations and comparisons.
3. Look for trends, patterns, or anomalies in the data.
4. Consider both short-term and long-term implications of the findings.
When preparing your report, adhere to these guidelines:
1. Be concise yet comprehensive.
2. Use clear, professional language.
3. Support your analysis with specific data points from the provided financial information.
4. Provide actionable insights and recommendations when appropriate.
5. Anticipate follow-up questions the CEO might have and address them proactively.
3. Agent Configuration
Key agent parameters include:
Agent Name
Agent Role - In AWS, everyone should have explicit permission to use the Intern's Laptop; the Agent role explicitly permits the Intern to use a company Laptop and company data.
The description gives you the ability to navigate through hundreds of Interns
FoundationModel - Dont ignore this parameter, it is the University that your Intern graduated from, it defines its intelligence, speed and cost. If you find some tasks that are less intelligent but required very frequently, you may hire other Interns via this field.
FinancialAnalysisAgent:
Type: AWS::Bedrock::Agent
Properties:
AgentName: !Ref AgentName
AgentResourceRoleArn: !GetAtt BedrockAgentRole.Arn
Description: Demo of CEO finance assistant
FoundationModel: arn:aws:bedrock:us-east-1::foundation-model/anthropic.claude-3-5-sonnet-20240620-v1:0
First day on the job on the Deployment Process
Before the Intern joins your team, I assume you already gave it permission to your well-defined and secure Data Lakehouse, defined tools to query your data and instructions to follow when asked questions.
Let's walk through the onboarding or deployment script:
- Build and Deploy Lambda Function
cd finance-tool-lambdasam buildsam deploy --stack-name ai-agent-finance-tool-$RANDOM_SUFFIX
- Deploy AI Agent Stack
aws cloudformation deploy \
--template-file bedrock-agent-stack.yaml \
--stack-name ceo-fin-report-$RANDOM_SUFFIX \
--parameter-overrides file://agent-params.json \
--capabilities CAPABILITY_IAM
- Prepare Agent for Use
aws bedrock-agent prepare-agent --agent-id $AGENT_ID
First question to the Intern or Testing Your AI Agent
For testing, Ill use a simple NodeJS script and AWS SDK for JS to ask the same question from the previous article: How did we do with the hospitality business in the first half of the year? Please send the report to email.
I assume that you have credentials via aws sso login and changed AgentId to yours. You can find AgentId in the newly deployed Cloudformation stack output or on the AWS Bedrock Agents page.
import {
BedrockAgentRuntimeClient,
InvokeAgentCommand,
} from "@aws-sdk/client-bedrock-agent-runtime";
/**
* @typedef {Object} ResponseBody
* @property {string} completion
*/
/**
* Invokes a Bedrock agent to run an inference using the input
* provided in the request body.
*
* @param {string} prompt - The prompt that you want the Agent to complete.
* @param {string} sessionId - An arbitrary identifier for the session.
*/
export const invokeBedrockAgent = async (prompt, sessionId) => {
const client = new BedrockAgentRuntimeClient({ region: "us-east-1" });
// const client = new BedrockAgentRuntimeClient({
// region: "us-east-1",
// credentials: {
// accessKeyId: "accessKeyId", // permission to invoke agent
// secretAccessKey: "accessKeySecret",
// },
// });
const agentId = "AJBHXXILZN";
const agentAliasId = "TSTALIASID";
const command = new InvokeAgentCommand({
agentId,
agentAliasId,
sessionId,
inputText: prompt,
});
try {
let completion = "";
const response = await client.send(command);
if (response.completion === undefined) {
throw new Error("Completion is undefined");
}
for await (const chunkEvent of response.completion) {
const chunk = chunkEvent.chunk;
const decodedResponse = new TextDecoder("utf-8").decode(chunk.bytes);
completion += decodedResponse;
}
return { sessionId: sessionId, completion };
} catch (err) {
console.error(err);
}
};
// Call function if run directly
import { fileURLToPath } from "node:url";
if (process.argv[1] === fileURLToPath(import.meta.url)) {
const result = await invokeBedrockAgent("How did we do with the hospitality business in the first half of the year? Please send the report to email.", "123");
console.log(result);
}
Launching this script via node index.mjs I got a response:
{
sessionId: '123',
completion: 'We had a relatively stable performance in the hospitality business during the first half of the year, with a slight decrease from Q1 to Q2. Here are the key points:\n' +
'\n' +
'1. Total bookings for H1 2024 were $282,607.99, with total billings of $255,309.52.\n' +
'2. There was a small decline in performance from Q1 to Q2:\n' +
' - Bookings decreased by 1.51%\n' +
' - Billings decreased by 2.76%\n' +
'3. The average billing rate slightly decreased from $90.92 in Q1 to $89.64 in Q2.\n' +
'4. May was our strongest month, while June showed the weakest performance.\n' +
'\n' +
"I've sent a detailed report to your email with further analysis and recommendations for improving performance in the second half of the year."
}
Conclusion
While this implementation requires traditional development skills, the power lies in decoupling Tools from the Brain. This separation enables:
Reuse of tooling across multiple Interns
Creation of specialized Intern teams
Complex cross-industry analysis capabilities
Rapid iteration and improvement cycles
Focus on GitOps practices for your AI Agents while trusting your developers with the coding tasks. This approach creates a scalable foundation for unlimited AI workforce expansion.
The complete CloudFormation template and deployment scripts are available in the GitHub repository.
Top comments (0)