Member-only story
The Future of Agentic HPC: What If 10,000 Agents Were Your Cluster’s Users?
Your coding agent has written the analysis. Now you have to tell the supercomputer how to run it.
How many processors? How much memory? How long should it be allowed to run? Ask for too little and the job can die halfway through. Overshoot and you can wait longer for a slice of a machine other people need too. The code may be ready while you’re still working out what to put in the request.
I want to be able to leave that analysis with the agent and come back to results I can inspect, including what failed along the way. Getting there means handling the work between writing the code and producing an answer. Waiting for resources. Recovering from a failed run. Checking that all the samples made it into the result.
That’s where I see a much larger opportunity for high-performance computing (HPC), the shared supercomputers researchers use for work that needs more computing power than their own machines can provide. We could design the service around helping an agent carry an analysis through, with a researcher steering it. Think about how much of your day currently goes into getting the next run started.
