Emile Baizel

Emile Baizel co-developed the Amazon Nova Act MCP Server with Yilin Zhu, giving AI agents tools to navigate websites, fill forms and extract information while reusing browser sessions. His work spans blockchain application infrastructure, agents that invoke wallet operations, and browser workflows that check whether an intended change actually appeared.

Connecting blockchain to web applications

Emile began working with blockchain in 2018 through an Ethereum hackathon. As a Senior Blockchain Architect at AWS, including in Professional Services, he developed implementation guides for connecting web applications to blockchain systems: authenticating users, submitting signed transactions and managing the credentials those operations require.

His serverless blockchain architecture replaced an EC2-hosted REST API with API Gateway and Lambda. A donation-tracking application illustrated the mechanism: ordinary web requests became signed Hyperledger Fabric transactions, with enrollment credentials stored in Secrets Manager.

His subsequent blockchain identity integration addressed a separate problem. Signing into a web application does not itself give a user the credentials needed to act on a blockchain. Emile connected Cognito authentication to Managed Blockchain identities, showing developers how to bridge those two systems.

He also created The Next Block, an interview podcast about people building blockchain products and projects, including developer tooling and decentralized infrastructure.

Giving agents transaction capabilities

In March 2025, Emile and Zhiyu Hao published a crypto-agent blueprint using Amazon Bedrock. A supervisor coordinates specialized agents that retrieve news, query blockchain data and invoke wallet operations. Lambda action groups implement the concrete operations, while AWS KMS manages wallet keys.

The design distinguishes what a model can decide from what an application should permit. Emile and Zhiyu identify transaction limits and approval workflows for high-value transfers as additional application-level controls. These are proposed extensions to the example, not protections it already guarantees. That distinction matters when a language-model response can lead to a financial transaction.

Emile also wrote an Amazon Bedrock integration for elizaOS, giving agents in that ecosystem access to Bedrock’s foundation models.

Browser sessions and visual feedback

Emile and Yilin Zhu released the Nova Act MCP Server together. It exposes browser capabilities through a shared interface for MCP-compatible agents, so developers can connect an existing agent without writing a separate Nova Act integration for each client.

Two decisions explain how the server handles web work:

• Fetch before opening a browser. The server first tries a lightweight URL fetch. When a task requires browser interaction, it preserves the session across subsequent actions, allowing the agent to continue working in the same browser context.
• Verify after acting. Emile’s practical guide recommends giving an action a focused objective, then inspecting the resulting page or extracting its data. A completed click establishes that an interaction occurred; checking the page establishes whether it produced the intended result.

In Build with Perception Agents, presented with Shruti Arora under an Amazon AGI Lab affiliation, Emile explored a complementary feedback loop. Visual annotation lets a person select an element, draw on a page or drop a pin, then export deterministic JSON feedback for a coding agent. Visual verification checks the rendered site against design expectations and user flows.

An audience demonstration made the distinction tangible: a coding agent reported adding a duck icon to a podcast page, and a browser verification pass checked whether the icon appeared. The example connects the two parts of the workflow—communicating a requested change precisely and inspecting the rendered result—without treating the coding agent’s completion message as sufficient evidence.