Inside the TPC Hackathon: How Researchers Are Building Agentic AI for Supercomputing

Thirty researchers from European and US supercomputing centers spent early September holed up at the Leibniz Supercomputing Centre (LRZ) in Garching, working through the messy practical problems of getting AI agents to actually do useful work on HPC systems. The event was the Trillion Parameter Consortium (TPC) Europe's autumn hackathon, and the output wasn't slideware. Seven working groups walked away with working prototypes, expanded codebases, and, maybe more importantly, a clearer sense of who else in Europe and the US is wrestling with the same problems.

That last part matters more than it sounds. Agentic AI, where large language models coordinate jobs and interact with other tools to carry out multi-step tasks, is still early enough that most teams are solving the same integration headaches independently. A hackathon that puts them in the same room for a week is basically a deduplication exercise for research effort.

TPC itself is young. It launched at SC23, the supercomputing conference, backed by Argonne National Laboratory (ANL), Japan's RIKEN Centre, and Spain's Barcelona Supercomputing Centre (BSC). The pitch, according to Prof. Charles Catlett, senior computer scientist at ANL and a member of the TPC strategy team, is to build "a sustainable movement for the use of AI in science" alongside research into making these systems trustworthy. TPC Europe followed in 2024, and got its own office and organizational backbone this spring when the EuroHPC Joint Undertaking funded the EuroTPC Project. LRZ has since taken point on drafting a EuroTPC roadmap, gathering input from researchers across the continent to shape EU AI-for-science strategy.

The consortium runs four hackathons a year; two land in Europe. This one followed a spring event at CINECA in Italy. The cadence is deliberate: regular touchpoints, not one-off conferences, seem to be the model TPC is betting on for keeping momentum between the big annual gatherings (the last one drew European members to LRZ itself, with sessions streamed between the US and satellite sites).

What Teams Actually Built

The technical range on display was wide, but a few threads stood out.

• One team worked on agentic workflows for hypothesis generation in protein research, essentially using LLM-coordinated pipelines to propose testable ideas rather than just crunch existing data.
• A group from the Gauss Supercomputing Centre (GCS) built early-stage workflows letting an AI agent adapt code across different supercomputer architectures, a real pain point given how much porting effort HPC teams burn just getting code to run on a new system.
• A separate GCS team tackled the security side: authentication and authorization workflows for AI agents, motivated by recent incidents where agent-based systems have overstepped their intended boundaries. If agents are going to run semi-autonomously on shared HPC infrastructure, someone has to define what they're allowed to touch.
• A Munich Centre for Machine Learning (MCML) team looked at whether language models can help researchers formulate and iterate on benchmarking algorithms faster.
• Another group focused on AI-assisted pandemic preparedness and risk prevention.

"For me, the most important thing during the TPC hackathons is the exchange of ideas," said Mitja Sainio, a machine learning engineer at Finland's CSC – IT for Science Centre, who worked with the GCS team on the code-adaptation project. That kind of cross-pollination is arguably the actual product of these events, more than any single prototype.

Ajay Navilarekal, an LRZ researcher who helped organize the sessions, noted that the seven groups came in with different goals. Some wanted a rough first prototype to keep iterating on solo. Others showed up with existing code they wanted to optimize and extend with fresh input. Both are legitimate hackathon outcomes, and it's a useful reminder that "hackathon output" doesn't have to mean a finished demo.

Energy efficiency got its own attention too. Navilarekal flagged the obvious tension directly: "The more agentic AI is used, the more energy is consumed due to inference workloads. We need to limit that." That's the uncomfortable flip side of agentic systems, more autonomy generally means more inference calls, more tool invocations, more compute burned per task. As agent frameworks proliferate across research computing, the energy accounting problem doesn't go away just because the workflows are novel.

Dr. Nicolay Hammer, who heads LRZ's Big Data & AI (BDAI) team, framed the broader push in terms of the EuroTPC roadmap effort: the center is actively collecting recommendations from researchers and institutes to shape EU-level AI strategy for science. Agentic AI, unsurprisingly, is set to be a central theme there.

Key Takeaways

The security angle deserves more attention going forward. GCS's work on authentication and authorization for AI agents is early, but it's addressing a gap that's going to widen as more labs deploy semi-autonomous agents on shared HPC clusters. Access control for a human user is a solved problem. Access control for an LLM-driven agent that can spawn sub-tasks and call external tools is not, and getting it wrong on shared research infrastructure has real consequences.

The cross-domain networking effect is the other thing worth flagging. Dr. Miguel Vazquez from BSC and Dr. Arvind Ramanathan from ANL, both working on AI for bioinformatics and genomics, said they left with new collaborators and project ideas rather than finished code. "The TPC community brings together a unique pool of expertise," Vazquez said. "This will enable us to successfully establish AI and agentic AI across all areas of research." Catlett, meanwhile, pushed the group to think beyond lab science entirely, floating the idea of AI agents analyzing urban camera and sensor data to improve city life.

None of this is glamorous in the way a big model release is. But it's the kind of infrastructure-level, community-building work that determines whether agentic AI in scientific computing becomes durable practice or stays a collection of one-off demos. Worth watching whether the EuroTPC roadmap, due out of this LRZ-led effort, turns these hackathon prototypes into something institutions actually standardize on.