[FEATURE] Remote Agent

[ylwu-amzn]

We propose adding a remote-agent capability to the ML Commons plugin, similar to the existing remote-model support. This will allow us to hook into powerful external agent frameworks.

Open‑source agent frameworks like LangChain, Strands (AWS) are advancing quickly. They bring built-in orchestration, memory management, multi-agent workflows, comprehensive tool integrations, and production-grade observability.

Our implementation approach mirrors the successful remote-model connector pattern. We plan to develop a unified interface that standardizes agent interactions, implements robust security measures, and provides a consistent data layer for agent operations. This architecture will allow users to run both local and remote agents interchangeably, providing flexibility while maintaining consistent security and operational standards.

This enhancement positions us to take advantage of current and future developments in agent technologies while ensuring a seamless experience for our users

[brianf-aws]

I feel that the challenge will be on having this new API be restricted on complying with existing local agents. If we think about remote models via connectors were done they allow for any kind of input and output and aren't restricted to the input that current for example text embedding models have.

POST /_plugins/_ml/_predict/text_embedding/cleMb4kBJ1eYAeTMFFg4
{
  "text_docs":[ "today is sunny"],
  "return_number": true,
  "target_response": ["sentence_embedding"]
}

• source

Here you can see there is specific parameters for text_embedding but currently user has option of using remote text embedding model with whatever parameters they want to provide.

Additionally I'm curious to know the reason behind compatibility with local agents. I'm assuming for good practice and for a multi-agent system? i.e. (OpenSearch Agent + remote Agent)