AWS re:Invent 2025 - Adopting AI within Streaming Architectures (AIM265)

Adopting AI within Streaming Architectures: Insights from AWS re:Invent 2025

Overview of Red Panda and Streaming Data

• Red Panda is a company that specializes in streaming data and data movement within organizations, operating on top of the Kafka protocol.
• They help organizations like the New York Stock Exchange and retailers move large volumes of data (tens of gigabytes per second) throughout their systems.
• Red Panda's platform can be deployed on-premises, in the cloud, or in hybrid environments, providing a single binary that is cloud-agnostic.
• The platform ingests data from various sources like databases, CDC systems, and microservices, stores it in an immutable log, and can sync data to common analytics solutions like Snowflake and Databricks.

Integrating AI and Agents within Streaming Architectures

• Red Panda has released an "Agentic Data Plane" that integrates AI and agent-based systems into their streaming platform.
• Agents are defined as autonomous entities with intent, access to language models, and a set of tools to perform complex tasks.
• The Agentic Data Plane runs within a private VPC or on-premises, providing access to various language models (e.g., Anthropic, OpenAI, Bedrock) without locking customers into a specific vendor.
• All agent interactions and outputs are logged to an observability layer, enabling replay, auditing, and governance.

Key Challenges and Principles for Adopting AI Agents

• Governance, security, and access control are critical challenges when integrating AI agents into an organization.
• The principle of "extending trust" is key - starting with a controlled environment to observe agent behavior before expanding access and usage.
• Red Panda's approach is to provide a layered architecture with an "Action Layer" for agents to operate, a "Data Plane" to log and govern all agent activities, and a "Reefing Layer" to control access to language models and other tools.

Leveraging Streaming Architectures for AI Agent Deployment

• Streaming architectures are already ubiquitous across industries, making them a natural fit for integrating AI agents.
• Streaming principles like publish-subscribe, asynchronous operations, and decoupled services can be applied to agent deployments, providing a familiar framework for organizations.
• Streaming also enables real-time context delivery to agents, which is a powerful feature for decision-making and event-driven architectures.

Practical Guidance for Adopting AI Agents

• Start by identifying and securing access to the data sources that agents will need, working closely with infrastructure and security teams.
• Deploy agents within secure VPCs and private networking initially, logging all activities for auditing and governance.
• Begin with well-understood, documented processes and use cases (e.g., manufacturing, IoT) to validate agent behavior before expanding to more complex scenarios.
• Focus on low-latency use cases where agents can provide immediate value, such as in adtech, manufacturing, and gaming.

Example Use Case: AI Agents in Manufacturing

• Agents can be deployed at the edge, near manufacturing equipment, to augment technicians and guide them through repair processes based on real-time data and contextual information.
• Agents can also be deployed in the operational database layer to provide predictive maintenance recommendations, leveraging historical data and spare parts information to identify potential issues before they occur.