AWS re:Invent 2025 - Building Scalable, Self-Orchestrating AI Workflows with A2A and MCP (DEV415)

Building Scalable, Self-Orchestrating AI Workflows with A2A and MCP

Overview

• Presentation on building autonomous, multi-agent systems using the Agent-to-Agent (A2A) protocol and Model Context Protocol (MCP)
• Focuses on designing self-orchestrating workflows that can adapt to non-deterministic AI outputs without central control
• Highlights the limitations of traditional orchestration approaches when dealing with generative AI agents

Key Concepts

Agent-to-Agent (A2A) Protocol

• Provides a consistent way for agents to discover each other and collaborate at runtime
• Agents publish "agent cards" that define their capabilities, input/output schemas, and versioning
• Allows agents to dynamically invoke each other's capabilities without hard-coding dependencies

Model Context Protocol (MCP)

• Standardizes how agents interact with tools and access context/data
• Defines strict input/output schemas for tools, ensuring predictable and reliable data exchange
• Removes ambiguity in model reasoning by working with structured, validated information

Agent Loop

• Execution model that every agent follows, regardless of underlying compute service
• Consists of four phases: Compose, Query, Execute, and Continue
• Ensures agents behave predictably, are re-entrant, and can safely run in a distributed system

Supervisors and Workers

• Supervisors handle global reasoning, dependency resolution, and workflow orchestration
• Workers focus on fast, deterministic execution of specific tasks
• Different model capabilities are selected for supervisors (more reasoning) vs. workers (more schema-focused)

Technical Details

• Agents are implemented as stateless Lambda functions, leveraging Bedrock Agent Core for memory and observability
• Momento cache is used for low-latency access to shared state and MCP tool responses
• Determinism is enforced through structured prompts (using the RISEN framework), item potency, and atomic API calls
• Observability is provided through Bedrock's dashboards and A2A's tracing capabilities

Business Impact

• Enables building scalable, adaptive AI-powered workflows that can handle non-deterministic model outputs
• Avoids the pitfalls of centralized orchestration by distributing control and decision-making to autonomous agents
• Improves reliability, reproducibility, and debuggability of complex AI-driven business processes
• Allows rapid experimentation and evolution of AI-powered systems without risking catastrophic failures

Example Use Case: Delivery Exception Handling

• Delivery driver reports that a customer was not home to receive a package
• A single agent handles the basic workflow of rescheduling the delivery
• In a more complex scenario, a package is damaged during transit
  • Triage agent assesses the situation and delegates to specialized agents (payment, warehouse, order)
  • Agents collaborate to refund the customer, replace the item, and update the order status

Key Takeaways

• Agents, not workflows, are the driving force behind this architecture
• A2A and MCP enable dynamic, deterministic collaboration between autonomous agents
• Supervisors and workers with different model capabilities are key to managing complexity
• Observability and determinism are critical for building reliable, debuggable AI-powered systems
• This approach can be applied to a wide range of AI-driven business processes and use cases