Apache Flink Agents

Apache Flink Agents is an official Apache Flink subproject, positioned as an event-driven streaming Agentic AI framework. It addresses the limitations of traditional AI agent frameworks in real-time streaming scenarios, providing millisecond-level response, large-scale distributed execution, and exactly-once action consistency guarantees.

Core Capabilities#

• Streaming Engine: Real-time event processing powered by Flink's distributed streaming engine with millisecond-level response
• Agent Paradigms: Workflow Agent (DAG orchestration) and ReAct Agent (reasoning-action loop)
• Consistency Guarantee: Exactly-once action consistency via checkpoint + external WAL
• Multi-language Support: Native Python and Java APIs
• DataStream Integration: Bidirectional conversion with Flink DataStream via from_datastream / to_datastream

AI Ecosystem Integration#

• LLM Providers: Anthropic, AzureAI, Ollama, OpenAI
• Vector Stores: Elasticsearch
• MCP Protocol: Model Context Protocol tool integration
• Memory Mechanisms: Sensory memory, short-term memory, long-term memory

Typical Use Cases#

• Real-time comment/feedback analysis: Extract sentiment, ratings, and pain points from streaming data
• Real-time monitoring and alerting: Intelligent anomaly detection and response based on event streams
• Multi-agent collaborative workflows: Orchestrate multi-Agent DAGs within Flink ecosystem
• Action-critical applications: Financial trading, inventory management requiring strong consistency

Runtime Requirements#

Development Example#

# Create execution environment and register resources
env = AgentsExecutionEnvironment.get_execution_environment()
env.register_chat_model("my_llm", chat_model_setup)

# Define Agent
@agent(prompt="...", tools=[...], chat_model_setup="my_llm")
class MyAgent:
    @action(input_event=InputEvent, output_event=OutputEvent)
    def process(self, event):
        return OutputEvent(...)

# Integrate with DataStream
stream = env.from_datastream(input_stream)
result = stream.apply(MyAgent)
result.to_datastream()

Observability#

Event-centric orchestration design where all agent actions are connected and controlled via events, enabling behavior understanding through event logs with monitoring and audit support.

Version Status#

Currently at v0.3-SNAPSHOT stage, under active development. Apache-2.0 licensed.