Why Events? Foundations
4 entriesCore ideas behind event-driven design.
Event-Driven Agent Patterns
10 entriesPatterns for coordinating agents and keeping long workflows reliable.
Agent Protocols & Event Standards
5 entriesStandards for how agents, tools and events describe themselves.
Brokers, Streams & Schemas
13 entriesInfrastructure that carries and stores events.
Event-Driven Agent Runtimes & Durable Execution
12 entriesFrameworks and runtimes for event-triggered, long-running agents.
Related Technologies on This Site
19 entriesEvent-driven, messaging and agent pages on this site.
Frequently Asked Questions
What is event-driven architecture for agentic AI?
It is a design in which AI agents communicate and are triggered through events, such as messages on a broker or stream, instead of calling each other directly. Agents publish what they did or learned, and other agents or services react to the events they care about.
Why use events instead of direct calls between agents?
Events decouple agents from each other, let many consumers react to the same fact, and, with a durable log, allow replay and audit. The sources in this directory also note costs: ordering, schema management and added latency, and several of them come from vendors that sell streaming platforms.
Which patterns are used?
Confluent describes four: orchestrator-worker, hierarchical, blackboard and market-based. Other guides from Anthropic, Microsoft, AWS and Google cover routing, concurrent and handoff patterns, and classic patterns such as saga, outbox and event sourcing help with long-running work.
What technologies are involved?
Event brokers and streams such as Kafka, Pulsar, NATS and the cloud event services, stream processors such as Flink, and runtimes such as Flink Agents, Solace Agent Mesh and Dapr Agents. Durable execution platforms such as Temporal and Restate keep long agent workflows running through failures.
How do A2A and MCP fit in?
MCP connects an agent to tools and data, and A2A lets agents from different vendors talk to each other. Event streams can carry the messages and state around both. Treat claims that a broker should sit under them as one architectural view, not a requirement of either protocol.