Sign in Subscribe
AI

How AWS re:Invent 2024 Sparked My Journey Into Agentic Software

Agentic Software was one of the AWS re:Invent 2024 topics that most inspired me. Find out why in this blog post.

Chris Ebert

• 6 min read
A A picture of Chris Ebert sending in front of a large AWS logo at the AWS re:Invent 2024 conference
Chris at re:Invent 2024

AWS re:Invent 2024 Recap

I had the privilege of attending the AWS re:Invent 2024 conference in Las Vegas this past December. Across five packed days, I immersed myself in sessions and hands-on workshops covering a wide range of topics, including:

As in previous years, AWS delivered several inspiring keynotes. Matt Garman's opening keynote energetically set the stage for the conference, but Werner Vogels' keynote resonated with me the most—as is typically the case. If you only have time to watch one keynote, I recommend watching Werner's insightful talk on the warning signs of complexity, "simplexity," and an introduction to Amazon DSQL: AWS re:Invent 2024 - Dr. Werner Vogels Keynote.

Out of everything I experienced, I left the conference most excited about agentic software's transformative and disruptive potential. This emerging trend, which enables the creation of intelligent task-oriented AI agents, is what I'm most eager to learn about throughout 2025. I will blog more on this topic in the coming months.

A Workshop that Made an Impact

The workshop "WPS401: Enabling Public Sector Efficiency through Generative AI" particularly resonated with me. This year, I noticed several workshops catered to specific business sectors, making the content more relevant and applicable to attendees. The WPS401 workshop presented a compelling scenario for using Amazon Bedrock to improve services for the City of Toronto. We explored creating an AI-powered service to assist citizens with accessing city parks. Key aspects of the solution included:

  • Using a large language model (LLM): Parsing text questions about parks by responding with information available in a park knowledge base.
  • Developing an AI Agent: Handling user park reservation requests by calling the appropriate APIs on their behalf.
  • Adding Guardrails: Preventing misuse and undesirable responses to ensure a safe, secure, and user-friendly experience by leveraging Amazon Bedrock Guardrails.

This experience was incredibly impactful because it directly aligned with my 14+ years of public sector work, making the use case both relatable and practical. Additionally, the workshop demonstrated how approachable building agentic software can be to all builders. Amazon Bedrock simplifies the development of agentic software on top of foundation models (FMs), making this cutting-edge technology accessible to a wide range of individuals—not just AI/ML experts. Furthermore, I was impressed by how quickly builders can create agentic software. We built a capable and practical agentic application within a two-hour workshop session.

In this blog post, I'll provide an overview of agentic software and why it emerged as one of my most significant takeaways from re:Invent 2024. Then, in a future follow-up post, I'll share how I built an example agentic application using Amazon Bedrock to create an AI agent that interacts with the U.S. National Parks API, demonstrating the potential and accessibility of building agentic software (you can preview that project here).

What is Agentic Software?

By now, most technologists have used an LLM such as Anthropic Claude, Google Gemini, or OpenAI ChatGPT. These tools are built on FMs trained on vast datasets and generate predictive text for diverse use cases. While LLMs are incredibly valuable as standalone tools, using them in innovative ways can unlock even greater potential.

"Agentic software" or "agentic systems" harness FMs to create autonomous systems that understand user intent and perform complex tasks. Athropic's Building Effective Agents blog post suggests further categorizing these systems:

  • Workflows: LLMs follow predefined code paths.
  • Agents: LLMs autonomously direct their own behavior to accomplish tasks.

This common language can be helpful when discussing agentic software.

With agentic software, we can build systems that understand user requests and autonomously determine the best course of action to respond. For example, consider an AI agent designed to handle HR requests for a large organization. For simple policy questions, such as questions about vacation policy, the agent can leverage retrieval augmented generation (RAG) to reference uploaded HR documentation as ground truth. Conversely, the agent can determine that querying an internal vacation REST API is the best approach for transactional requests, like requesting a vacation day or checking balances. The HR department may be concerned that the agent could potentially disclose undesirable information, and it can prevent such scenarios by adding guardrails to control its responses.

Agentic software can combine existing data, services, and even other AI agents in innovative and powerful ways. The potential for agentic software is clearly evident to me. Additionally, frameworks like Bedrock's AI Agent make building such software approachable to a broad range of builders.

Agentic Software Development is Accessible

Before re:Invent 2024, I hadn't given much consideration to Amazon Bedrock or shown interest in using it. However, re:Invent changed my perspective by demonstrating the utility of this framework and AI frameworks in general.

Amazon Bedrock is a framework that offers several essential features for building AI applications. First, it provides a consistent interface for interacting with different FMs. Through Amazon's partnerships with various AI vendors, developers seamlessly switch between models like Anthropic's Claue or Meta's Llama 3 with simple configuration changes. This flexibility enables rapid experimentation and opens new possibilities for development.

Second, Amazon Bedrock offers robust debugging tools, such as traces, that help visualize an agent's reasoning process. Determining why an agent behaved a certain way can be confusing, so having access to its chain of thought is tremendously helpful. Again, Bedrock provides a consistent trace interface no matter which FM you choose for your application.

The Bedrock Framework offers intuitive APIs and constructs for building AI applications. Developers can easily define Action Groups that specify which actions agents can perform on behalf of users. These Action Groups can call functions directly or interact with APIs defined in standard OpenAPI 3.0 specifications. By chaining multiple Action Groups together, builders can create sophisticated agentic software with defined guardrails to ensure agents operate as intended.

Some capabilities, such as building knowledge base agents, have become commodity solutions in Amazon Bedrock. With a few clicks, any developer with little AWS experience can create a knowledge agent in the AWS console. I expect an increasing number of AI interactions to become commodities soon.

The comprehensive tooling offered by AI frameworks, such as Amazon Bedrock, makes agentic software approachable to a broad audience. You don't need to be an AI/ML expert to leverage FMs to build innovative features for end users. With the barrier to entry significantly lowered, developers have more opportunities to create impactful AI-powered applications. This lowered barrier to entry also makes building agentic software more feasible for competitors and market disrupters.

What Makes Me Interested in Agentic Software

I'm fascinated by the emerging potential of agentic software, which I believe will be an evolving area of innovation in 2025 and beyond. Agentic software applies to many scenarios, including my personal and professional life.

I'm particularly interested in applying this technology to a multi-tenant production application I've helped develop. This application's support workflow currently has a bottleneck: while we have a support control plane, many operations, such as modifying data ownership, require escalation from our L1/L2 support team to an engineering team member. While our engineers can handle these requests, this arguably isn't the best use of their time.

I envision creating an AI agent to allow support staff to directly request application support actions. Using Bedrock Agent Action Groups, we can give the agent access to our existing RESTful APIs and Python scripts. Then, the agent can use these actions to assist end-users with support tasks. This approach leverages the Amazon Bedrock AI Agent framework to create a support operations interface without building an entirely new application. This AI agent could reduce engineering time responding to support requests. Over time, certain operations could be exposed to end users directly, eliminating the need for them to call support for specific requests.

Conclusion

If you're a cloud builder who hasn't been following developments in agentic software, I encourage you to explore this space. Amazon, Anthropic, the re:Invent conference, and practitioners in this space offer excellent resources on this topic. Agentic software is positioned to transform how we build software in the future. The APIs we create today will likely be used by AI agents tomorrow, making clear documentation and specifications more critical than ever.

In my next blog post, I'll share my experience building an AI agent that helps users explore U.S. National Parks. I'll discuss things I learned along the way and hope to inspire other builders to experiment with creating agentic software.