Agents that cannot assume somebody else pays for the compute
Originally posted on LinkedIn.
Most of the agent research I read quietly assumes somebody else is paying for the compute. A large model behind an API, context to spare, a network that is always there. Take those three things away and most of what we have built stops working.
That gap is the entire point of PerAgents 2027, the first international workshop on pervasive agentic systems and multimodal foundation models, co-located with IEEE PerCom 2027 in Goa, India, from March 8 to 12, 2027.
I am serving on the technical program committee, and one of the listed topics sits right on top of my own work. Decentralized, split and federated agentic learning. The rest of the call covers compressed language and vision models, multimodal sensor fusion, bandwidth-efficient distributed inference, cross-domain agent adaptation, and the question I find hardest of all. What trust, safety and local override should look like once an agent is allowed to move something in the physical world.
If you work on any of that, the call for papers is open now.
Submissions close 3 November 2026, anywhere on Earth. Six pages plus one additional page, references included. Notifications 8 January 2027, camera-ready 2 February 2027.
The official flyer is below. Please pass it to anyone whose work belongs in that room.