Building an agent mesh in 3 days: Syngenta’s Bedrock AgentCore journey

A large multinational enterprise, Syngenta has over 30,000 employees in more than 90 countries. Many teams across various functions and business units were experimenting with AI, each running their own initiatives.

Syngenta wanted to create a company-wide, interoperable environment that teams could use to continue innovating freely with AI while implementing the right security and privacy controls. It also aimed to reduce technical debt. “We want to balance the autonomy to innovate within functional teams with enforcing guardrails at a central level,” says Krohn. “We also saw a lot of rebuilding of the same sort of AI agents. They couldn’t get reused easily, either because they weren’t built to a common technical standard, or because people couldn’t find them.”

To build a solution, Syngenta used Amazon Bedrock AgentCore. “Amazon Bedrock AgentCore is a low-friction solution because a lot of our technology already exists in AWS,” says Krohn. “The newest upgrades to the solution seemed very promising and align well with changes in the industry.”

Using Amazon Bedrock AgentCore, Syngenta built an agentic mesh solution with centralized governance and accessible spoke accounts that business units can use to develop and experiment with agentic AI. To build the solution, Syngenta participated in a 3-day AWS Experience-Based Acceleration (AWS EBA)—a hands-on, agile, and immersive engagement. During the AWS EBA, Syngenta and AWS set up the architecture for the central hub account and two business unit accounts. Using AWS expertise and the prebuilt patterns in Amazon Bedrock AgentCore, the company deployed rapidly. “The AWS EBA helped us create our blueprint architecture for both individual agent and product development, as well as the global environment,” says Krohn. “Out of that, we’ve built a software development kit that teams can use to build agents with the standards that we’ve chosen.”

The agent mesh uses a multi-account hub-and-spoke architecture. The central hub account provides orchestration, observability, and gateway services using the gateway capabilities in AgentCore, which provides an easy and secure way to build, deploy, discover, and connect to tools at scale. Other accounts are connected to the central account using the runtime in AgentCore, which provides a secure, serverless, and purpose-built hosting environment for deploying and running AI agents and tools. “By using Bedrock AgentCore Gateway to deploy Bedrock AgentCore Runtime into cloud accounts at the individual function level, we give teams the freedom to deploy and run agents independently while maintaining central orchestration,” says Krohn.

The agent mesh uses Agent2Agent and AG-UI protocols for communication between agents, as well as Model Context Protocol for data retrieval. Syngenta handles identity management using AWS Identity and Access Management (AWS IAM) to securely manage identities and access to AWS services and resources and identity management in AgentCore for managing agent identities and securing credentials.

The solution also uses catalogs and registries to let business units find and use agents that other teams have already built. Now that Syngenta established a common technical standard, any agent can talk to the others within the agentic mesh.

Syngenta’s agent mesh was operational in 3 days during the AWS EBA. The company’s supply chain and marketing business units are actively deploying agents, with marketing building six agents in 4 weeks. Three additional functions—with use cases such as agents for cash management and HR ticket automation—will launch in the coming months. Using the runtime in AgentCore, Syngenta achieves 20 percent better latency than using physical servers, and deployment takes around 40 percent less time.

One widely adopted agent solution, called Austyn, that was implemented during the AWS EBA uses the code interpreter in AgentCore—which lets AI agents write and execute code securely in sandbox environments—to perform data analytics. Now up to 12,000 internal monthly active users bring their data to Austyn to run exploratory data analysis, plot charts, and create visualizations.

Going forward, Syngenta will refactor some agents that were built before the agent mesh to incorporate them into the solution as it continues to build new use cases. “AI should fundamentally change how we do business,” says Krohn. “At the end of the day, it will mean better products to help feed an ever-growing world.”