Planning Flink Agents 0.2

[xintongsong]

Time Plan

• Target Feature Freeze Date: Jan 15
• If the major features are completed much earlier, we might also consider to release earlier.

Features

• Experience improvements based on user feedbacks
  • todo: go through and check the existing github issues
• Action Execution and Consistency - @Sxnan
  • Refactor current action execution and consistency implementation
  • Support fine-grained durable execution
  • Support async-execution in Java
  • Support object storage based action store
• Comprehensive Memory Management - @wenjin272
  • [Feature] Introduce Sensory Memory #310
  • Short-Term Memory: collection type supports capacity limit and auto archiving to long-term memory
  • Long-Term Memory: vector based, similarity search, compaction
  • Knowledge (optional)
• Java API completion
  • Provide more built-in integration implementations
    • [Feature] Azure LLM Integration for the Agent #286
  • Embedding Model
    • [Feature] Support ollama4j embedding models in java #304
  • Vector Store
  • MCP
  • Async Execution
• Cross-Language supports: support using Python/Java resource in Java/Python actions - @GreatEugenius
• Support multiple Flink versions

[tsaiggo]

Just had a thought—what if we started to templatize and structure our prompts? Something like Microsoft's POML (https://github.com/microsoft/poml).
The upside would be way better readability, plus we could finally version control them like normal files. The downside is I'm worried it might make the project a lot heavier or more complex.

It’s just a brainstorm, haha, but I'm curious about the potential value.

[xintongsong]

This is an interesting topic. I think prompt management is definitely an important feature and a good improvement for the project. But given that Flink Agents is still at an early stage and there are quite some essential functionalities missing, this is probably not a priority we should focus on at the moment. WDYT?

[tsaiggo]

Completely agree. We should focus on the more critical features right now. Let's plan to revisit this once the project is more mature.