Introducing AISAC: A General-Purpose Artificial Intelligence Scientific Assistant for Research Environments
Abstract: Scientific artificial intelligence (AI) is at an inflection point. Recent co-scientist systems have demonstrated that language learning model (LLM) agents can generate hypotheses, synthesize literature and execute experiments — but these systems are built for demonstration, not for sustained deployment inside scientific institutions. Real research environments impose requirements that go largely unaddressed: provenance and reproducibility of reasoning chains, transparency and auditability of agent behavior, safe and governed code and data execution, and deployment within institutional constraints including vetted inference endpoints, high performance computing (HPC) infrastructure and federated authentication.
AI Scientific Assistant Core (AISAC) is a production multi-agent framework developed at Argonne National Laboratory to address this gap. Rather than proposing new agent algorithms, AISAC contributes a governed execution substrate — a reusable infrastructure stack on which many domain-adapted scientific co-scientists can be built and deployed without requiring domain scientists to become framework developers. This talk details the architecture and design principles behind AISAC. We describe its bounded driver/helper orchestration model, which enforces strict role separation between planning and execution agents and applies hard limits on recursion depth, delegation steps and tool rounds.
We discuss context engineering as a first-class problem — dual-floor budget management, non-destructive message trimming and retrieval-augmented generation-aware escalation that preserves evidence fidelity rather than silently summarizing. We present AISAC’s layered persistence model spanning conversation memory, semantic recall, user preferences, per-agent notes and post-run skill learning — enabling cumulative usefulness without fine-tuning. We cover the live observability and runtime steerability model that makes agent behavior inspectable, replayable and interruptible. Finally, we describe deployment across laptop, HPC and air-gapped environments from a single codebase, and interoperability with model context protocol partner agents, skills and standard industry schemas.
AISAC is currently in active use across six scientific domains at Argonne including combustion science, materials research, critical minerals and energy process safety. We close with lessons learned from production deployment and the path toward broader Department of Energy adoption.
See all upcoming talks at: https://www.anl.gov/mcs/lans-seminars.