🦄 Making great presentations more accessible.
This project aims to enhances multilingual accessibility and discoverability while maintaining the integrity of original content. Detailed transcriptions and keyframes preserve the nuances and technical insights that make each session compelling.
Overview
📖 AWS re:Invent 2025 - Autodesk Assistant: Building an Agentic Platform on Bedrock (ISV322)
In this video, Matthew Liem from Autodesk and Sahil Saini from AWS present Autodesk Assistant, an agentic AI platform built on AWS Bedrock. They demonstrate how the assistant automates design workflows in Revit through natural language commands, from calculating window-to-wall ratios to modifying building elements and generating documentation. The architecture leverages MCP servers to bridge LLMs with Autodesk APIs, domain-specific agents for different products, and a Conversation Proxy Router for intelligent request routing. The platform uses AWS Agent Core services including runtime for secure compute, Gateway for API-to-MCP conversion, Memory for context management, and Identity for access control. Strands framework was adopted as the standard for building agents across Autodesk teams, enabling the shift from "point and click" to "describe and do" workflows that dramatically accelerate the design-to-production cycle.
; This article is entirely auto-generated while preserving the original presentation content as much as possible. Please note that there may be typos or inaccuracies.
Main Part
Introducing Autodesk Assistant: From Design Challenges to Agentic AI Solutions
Hello everyone, thank you for coming to our session. I'm really excited to be here today to talk to you about Autodesk Assistant and how we are building an agentic platform on top of AWS Bedrock. My name is Matthew Liem. I'm a software architect at Autodesk, and I'm joined here by Sahil Saini, who's a solution architect at AWS and someone who's been an incredible partner throughout this journey. I know we only have 20 minutes and a lot to go through, so let's dive right in.
For those of you not familiar with Autodesk, Autodesk is a global leader in 3D design, engineering, and entertainment software. We build the software that architects, engineers, designers, and creators use to build everything from cars to games to machines—essentially anything that needs to be imagined, designed, and built. Autodesk has been a pioneer in the design and make space for over 40 years now, and as our customers have grown, so have we, from a handful of innovators to over 15,000 people globally.
Our industry solutions and platform services allow our customers to take an idea and turn it into reality from design all the way through to manufacturing, production, construction, and beyond. Today our customers, like many of yours, are facing unprecedented amounts of disruption and demand. Manufacturers are feeling challenged to constantly push new products out to the market while dealing with supply chain delays and labor shortages. In our AECO industry, designers are constantly feeling pressured to deliver new products and more complex projects faster while dealing with the demand for more infrastructure, more housing, more experiences, all while being more sustainable. Production studios of all sizes are struggling to keep up with consumer demand, tight timelines, and the constant demand for fresh new content.
Agentic AI represents a fundamental shift in how AI systems function. In the design and make space, it is allowing us to move beyond the old model of point and click to this new era of describe and do, where your intent becomes your actions. It is also unlocking entirely new levels of creative velocity for our customers. The amount of time it will take to go from an idea to impact is being shortened to moments, no longer weeks, allowing our customers to really focus on innovation and spend less time on day-to-day busy work. Agentic AI is also redefining what's even possible for our customers, signaling the shift from manual tasks to highly connected, automated, and accessible workflows. All of this is allowing our customers to unlock faster, smarter outcomes from optimizing their site plans to streamlining design iterations to automating post-production sites.
That's why we built Autodesk Assistant. Autodesk Assistant is your agentic AI partner that speaks design and make. From day one, it's able to do three things: automate the manual, connect the disconnected, and deliver real-time insights all in collaboration with their human counterpart. Let's jump into a quick demo to see it in action.
Right here is Revit, our building information modeling software, and on the right-hand side here I have Autodesk Assistant. The first thing I'm going to ask the assistant to do is tell me more about the building model that I have opened up here. This is a very common tell me or ask me type use case, very similar to a software engineer entering a new code repository and wanting to quickly get up to speed. So the first thing you want to ask is to be able to ask questions about it. In this case, I'm going to ask it to tell me the window to wall ratio of the building on my north facade.
Autodesk Assistant is going to go in, do its calculation, and return back around 19 percent. The next thing I'm going to ask the assistant to do is more of a do it for me use case or a design with me use case. In this case, I'm going to ask the assistant to take all these windows—we didn't like the 19 percent—and modify it to double hung windows with slightly larger dimensions so that we're able to get a little
more light. Behind the scenes, Autodesk Assistant is taking this prompt and routing it to one of our AI agents, which has a number of MCP servers behind it. It's calling an LLM which helps decide which MCP server and tool to use, which in turn calls one of the product APIs. So the assistant is going to try different things, and just like that, all my windows were modified. This is really powerful, especially for users like me who are not experts at Revit, but I'm able to go in, use natural language, and make all these modifications.
The next thing I'm going to ask the assistant to do is help me with some of my workflow documentation work. Once a model is complete, there's a lot of manual work that goes into making it ready to be distributed to other contractors or architects to take the design forward. So the first thing I'm going to ask the assistant to do is create a floor plan view of my building. It's going to go in and create a new sheet of my floor plan. It's a little bit messy, so we need to clean it up.
Next, I'm going to ask the assistant to apply a view template on top of this and also help me annotate some of the doors, windows, and walls. I'm going to ask the assistant to do that and let it run. The assistant is going to go in, look at my current sheet, identify all those objects, and create a new view that has all those different annotations.
We're almost done. The last thing I'm going to ask the assistant to help me with is create a cover sheet with this view. A cover sheet has a lot of our company standards, the numbering, naming, and so on. That's the last step before we're able to share it with others. Just like that, I was able to go from taking an existing model, being able to ask questions about it, being able to modify it through natural language, and then get it into a state that is ready to be shared with others.
Architecture of Autodesk Assistant: MCP Servers, Agents, and the Conversation Proxy Router
Next, let's look at the architecture that powers all of this. Starting with number one, our MCP servers. Our MCPs are a core component of our Autodesk Assistant architecture. At Autodesk, we've been investing in data and APIs for the last decade, so MCPs is a natural extension of that journey. That allows us to bridge LLM models and our APIs to real-world actions. We have both local MCP servers as well as cloud MCP servers. The majority of our customers use Autodesk products through our desktop, but there's still a lot of integration with our cloud services, so that's why we support both.
Number two is our agents. We have a wide variety of agents at Autodesk. We have product-specific agents like an AutoCAD agent or a Revit agent. We also have more general agents that span across multiple products, such as a customer support agent or an agent that helps with administration and licensing. These agents essentially provide the core capabilities of Autodesk Assistant, everything from question and answers to helping automate some tasks and workflows. Within these agents, another important thing to call out is that there's a wide variety of complexity. Some agents are very simple and are simply just calling an LLM to help with MCP tool execution, while other agents are a bit more complex and have a multi-stage workflow within them.
Another core component of our architecture is our Assistant Backend. How we've built Autodesk Assistant is as a shared platform service across our whole organization, so any team or product can bring their agent or MCP server and build on top of our platform. As a platform team, we've been focusing a lot of effort on shared capabilities like guardrails, context management, and how we're handling observability of token usage and LLM tracing.
Another key point is the Conversation Proxy Router. This is essentially the brains of Autodesk Assistant because it is responsible for all the routing of requests across all our products to the correct domain agent. It also handles some of the agent-to-agent communication that we have here. We use a number of different strategies within this Conversation Proxy Router.
In some cases we're using an LLM to help with the decision. In other cases we have more deterministic flows where it's rule-based. So for a given product, it will always route to a specific agent. When a user submits a query, they're interacting with our Assistant UI. We have many different flavors of our Assistant UI. We have the embedded version that you saw in the demo, but we also have this in a web version. Regardless of where you're accessing the assistant, you're always going to get the same look and feel, and the overall general experience is going to be the same.
The Assistant UI plays a very important role in the sense that it passes the user prompt to our backend, but in addition to passing the user prompt, it also passes applicable context. This context can be who the user is, where the user is coming from, what product they're from, what license they have, and all the applicable MCP tools. This is really important because we use this context to help with the decision making for agent selection as well as MCP tool selection.
Once the request gets routed to one of the agents, the agents process that request and send it back to the Assistant UI to be exposed to the user. That is essentially our end-to-end flow of how we process requests through our Assistant architecture. I'm now going to pass it off to Sahil, who's going to go over Agent Core and some of the underlying services.
Building Production-Ready Agents with AWS Agent Core and Strands Framework
I'm Sahil, a Solution Architect with AWS specializing in generative AI and agentic AI. I work with strategic partners like Autodesk in building the next generation of generative AI applications. Autodesk was building this whole agentic workflow, and they realized at a very early stage that building these agents is great and good for business. However, taking it to production caught their attention around the performance, scalability, security, and the governance of agents and the workflows that they were building.
A platform team was building agents. They took a step back and tried to make it more like a platform strategy rather than building it as bespoke or snowflake agents. During the exploration phase, there are a couple of challenges that organizations like Autodesk figured out at an early stage. Specifically, they needed to understand how to get a secure compute runtime environment for agents as well as for their MCP servers. They needed centralized context management so that all the industries across Autodesk can have the same contextual information when the workflow goes between the agents.
They also needed centralized identity and access controls on top of the Autodesk Assistant, and they needed to discover and connect with custom MCP servers or cross-domain agents for any purpose-built workflows. Without saying it, auditability and observability are important callouts. Autodesk figured out at an early stage that these are the things they cannot let all the industry teams build by themselves. Why not have it as a platform strategy so that everybody can plug it in, use the best practices, and build their workflows on top of it?
That was the Autodesk platform approach. Autodesk built a centralized Autodesk Assistant, which is a frontend UI so that all the end users across different industries and personas will have the same consistent experience. They will not see a different UI. Behind the scenes, they will be talking to platform components that I'll be going over in the next slide, and all the domain agents or MCP servers will integrate to this platform to make sure that they get the best out of what has already been vetted.
Going next, this is a high-level overview of how Autodesk laid this across. Matt talked about how Autodesk built this centralized agent, and that's where Autodesk uses Agent Core runtime and uses Strands as a framework to power all the interactions or the workflow to the domain agents and MCP servers. Every time a user query comes in, this is a centralized orchestration agent powered by Strands that takes that call, whether that query needs to be solved by a default agent, a domain agent, or a support agent. All these agents have their own native integration to hybrid MCP servers. Moreover, Autodesk has doubled down on Agent Core Gateway because Autodesk has invested a lot on API, on data, and how to quickly convert those APIs into MCP servers.
That's where Autodesk uses Agent Code Gateway, where all the Autodesk APIs that are open and enabled are being directly exposed through Agent Code Gateway and exposed as MCP servers. Moreover, all the existing MCP servers that have been created or built have been interfaced behind Agent Code Gateway. If you look at the top, Autodesk, with a platform mindset, made a couple of capabilities as centralized platform capabilities coming from guardrails. Every conversation that is coming in or going out of the Autodesk Assistant has been vetted thoroughly with all the guardrail policies inferred at the time of inference.
Similarly, Autodesk leverages Agent Core Memory and AWS storage services to store all the short-term and long-term conversations so that all the agent workflows going across the domain have the same contextual information. Agent Core Identity has been leveraged to power all the MCP workflows and to make sure agents have centralized identity while interacting with the underlying downstream applications. Autodesk also built a hybrid cloud MCP server powered on AWS services for all the cloud MCP as well as the desktop MCP.
I know I talked a lot about Agent Core. Let me give you a quick overview of what Agent Core is for people who don't know about it. These are a set of fundamental services that Amazon released about six months ago, and it achieved general availability very recently. It helps customers build, scale, and roll out agents to production quickly. In Agent Core, the fundamental concept involves primitives, and you don't have to use all of them. You can use what is required for your agent protocol. Agent Core also offers open foundation model support. You can interact with models from Bedrock, from OpenAI, or whatever your model inference platform is, and you can use any framework of your choice.
I'll quickly touch on a couple of things that Autodesk uses and that has been used as a centralized platform, specifically around runtime. Autodesk figured out that many of the industry teams have their own preference for frameworks. A couple of teams use Langchain, a few teams are comfortable with CrewAI, and Strands is predominantly used by many teams. So Autodesk uses Agent Core runtime for secure, isolated compute infrastructure so that all the industry teams can plug into it and have a production-ready compute environment for running their MCP servers or their domain agents.
Agent Code Gateway is a service offering that lets you expose your open API spec, your existing MCP server, or any Lambda functions that you have built for your business workflow. It helps you integrate and gives you quick turnaround for MCP interfaces along with Agent Core Identity. Finally, on Agent Core Memory, to have centralized contextual management and to make sure that workflows have short-term information and all the previous conversations that happened with the Autodesk Assistant, Autodesk vetted on Agent Core Memory and uses all the short-term and long-term memory to power the workflow.
Going next, I want to quickly touch on Strands Agent because when many teams within Autodesk were using different frameworks, there was also a lack of opinion on what to go ahead with. That's where Autodesk made Strands the framework of choice so that all the industry teams while building their agentic framework can quickly get onboarded with native integration with Agent Core. A couple of features about Strands: it has support for all the open source protocols, it has support for MCP, it's easy to build, and Autodesk has seen greater adoption and seamless adoption of Strands. It accelerates the development cycle for all the industry teams to quickly build agents.
I know we have a couple of seconds left. I'll quickly go over the key takeaways. Autodesk realized that while building a system like concurrent sessions, they need a scalable runtime environment, and that's where they leveraged Agent Core runtime with other AWS services that they were using. For platform components, they unified the framework and made sure agents and MCP servers are common so that industry teams don't have to switch between different kinds of services and can get to the work.
At a high level, these are the key takeaways. I'll move to the call to action. Deep dive into Agent Core. We have a lot of samples there. We have also built what we built for the Autodesk Assistant as an agent, Omnimesh, which is an open source framework for multi-agent collaboration. We encourage you to test it in your AWS account, and finally, we encourage you to check the Autodesk Assistant in action to see all those cool features. With that, thank you very much. We'll be around here, so if you have any questions, I'm happy to answer.
; This article is entirely auto-generated using Amazon Bedrock.
Top comments (0)