AID2E Technical Meeting

Date: June 18, 2026

Phase I

• Derek
• all PRs merged in the development branch

• Functionality to support cross-communication between multistep jobs * In progress at: https://github.com/aid2e/AID2E-framework/pull/55

• Coming next: DTLZ as discussed in the last meeting

• (confirmed) For the paper - analysis/results we could start from BIC (we can use this and “duplicate” for stress-testing complexity)

  • Wait for dRICH and FF to be included

• Amit

• Multi-stage PanDA runner for BIC MOBO

• Projected Timeline (first preliminary results supporting paper)

• DTLZ – (cross-communication + CLIs work needed) ~ early July to see some first results

• BIC – (if we see something from DTLZ not far for BIC too; cf. multi-stage job; ) ~ mid July

• dRICH, FF → outline similar path forward (for the paper)

Phase 2

• (Amit) Amit gave a presentation at the last general meeting [LINK] and laid out challenges

• A few MCP servers for different task, maybe converged later

• Resources - performance with open model
• finish integration of MCP that helps analyzing the PanDA jobs/logs
• The other one is for the AID2E agent (assistance - steer optimization)
  • Steer — (Cris) comment
    • Optimization engines can be more effective in making certain decisions inherently during the optimization process
  • Let’s define assistance…
    • At a bare minimum, can we define assistance as supporting the definition of the optimization problem and get the user started with launching AID2E jobs?
    • (James) Two aspects
      • Physics — figures of merit / objectives
      • Code-side
  • Passive/active agentic workflows
    • Monitor
    • Code/Optimization Assistance

• (Amit) Build Closure tests

  • Assistance Closure test 1 - e.g., building a working optimization pipeline for the AID2E assistant
  • Monitoring Closure test 2 - e.g., you know the issue, is it found (flag raised)? Being able to interact with PanDA server;
  • (James) one issue is going to be scalability at some point (not ingest the entire thing) - regex match

• Wen worked on LangChain and LangGraph

• (Cris) Let’s start from one of the challenges; what data do we have?

• Discussion on Logging (PanDA vs SLURM)
  • Do we need to make them accessible to all users?
  • PanDA - possible to do that for AID2E
  • SLURM -
  • We can filter out a lot of information that is not really needed for monitoring - how do you do this during execution?
  • Is someone working to replace the current PanDA monitor with an LLM-based one?
• Choice of LLM
  • Do we have enough resources to run the LLM? In the future for bigger tasks?
  • Quantify ‘how large’ LM are needed for this

Action items

• (Amit) volunteer on creating a document with a list of things that we need to support all this [LINK]