ALL NPPS Members

[Torre Wenaus <wenaus AT gmail.com>]

Hi all,

Hong is putting together a revised AI/ML FWP based on DOE HEP feedback, he wants an agentic AI section based on the input we've given on MCP application. This is what I've provided. Comments welcome...

it's also in a googledoc

A new and rapidly growing AI domain, as significant as the arrival of LLM technology itself, is active agents representing services that can be interrogated and used by LLMs and other intelligent interfaces to create assistants/actors with powerful capability to act in real time, based on current information direct from and curated by the services themselves. No hallucinations from bad data (the data provided through agents is curated by the services), no blindness to the time since training, live real-time knowledge, and no limits (but complete control) on the information and capability an agent can provide for its services.
The basis of this is an open source protocol established recently by Anthropic, the Model Context Protocol (MCP). MCP agents leverage the APIs provided by the services, and they are an effective driver to improving and thoroughly documenting those service APIs, benefiting the entire project. Effort invested in MCP agents is effort invested in our own systems to integrate them with the emerging ‘agentic ecosystem’. The speed and thoroughness with which the AI and software community has adopted MCP gives confidence it is an approach we can safely adopt and evolve with.  

MCP is a technology that can be the real basis for "end to end integrated AI from collider to detector to analysis" for EIC, for greatly increasing intelligent automation and user-facing capability in the systems and services we develop, like the PanDA workload manager and the Rucio data manager. It is oriented more towards facility/ops than research applications. It is an enabler for autonomous control of accelerators and experiments. It is shovel-ready, with promising prototypes already underway in PanDA and EIC streaming readout, and many more prospects for early application including the ops environments of running experiments, where it could be a great aid to making operations and shifts better equipped with intelligent analytics and less demanding of effort. It is equally promising in the offline software and analysis environment as a live, domain-aware assistant in using experiment software and applying it in analysis; it can be a strong enabler for the user-centric software design approach adopted by next-generation experiments such as FCC and ePIC.
MCP agent development is work that can deliver substantial short-term return from moderate investment that integrates well with existing services development, improving the overall quality, capability and documentation of software services through the discipline of developing MCP APIs that most effectively represent the service to LLMs.

We propose to grow our present exploratory R&D into an MCP agent development laboratory, developing MCP services for the many software services and systems in which we hold core expertise, with prioritization informed by where these developments can have the greatest impact for our experimental programs. Complementing MCP development itself we will build an infrastructure providing agentic AI environments to experiments consisting of domain-aware, optimized LLMs integrating our MCP services.

  Torre 

--

-- Torre Wenaus, BNL NPPS Group Leader, ATLAS and ePIC experiments

-- BNL 510A 1-222 | 631-681-7892