Research

Research in Agentic Computing.

We study the systems abstractions and mechanisms required to efficiently, reliably, and scalably support autonomous agent workloads.

Our thesis is that agent-native workloads introduce execution structures, resource dynamics, and runtime behaviors that differ from conventional applications and static AI inference workloads. Existing systems mechanisms therefore need to be re-examined, adapted, or redesigned.

Research Agenda

Three connected research threads.

01

Agentic Execution Models

How dynamically formed agent workflows, subagents, dependencies, communication, and state should be represented and controlled.

Focus

We study execution semantics for workloads whose control flow can emerge at runtime: spawning, delegation, communication, state capture, replay, verification, and guarded execution. Current evidence includes Agentivium Core contracts and HPC-Claw as a concrete domain system for verifiable decisions.

Dynamic execution graphsSubagent spawningExecution semanticsState and replayVerificationGuarded execution
02

Agentic Runtime & Resource Management

How models, compute, memory, network, tools, and budgets should be allocated as agent workloads evolve under uncertainty.

Focus

We investigate runtime prediction, scheduling, uncertainty-aware adaptation, multi-objective optimization, and compute-aware orchestration. HPC provides a realistic resource-constrained environment and a source of scheduling mechanisms, while the research questions extend beyond HPC.

Runtime predictionSchedulingUncertainty-aware adaptationMORLMulti-objective optimizationCompute-aware orchestrationValue-per-compute
03

Agentic Memory & Inference Systems

How workflow structure changes context, memory, KV-cache reuse, model serving, and routing for agent workloads.

Focus

We study whether workflow structure can reduce redundant inference through prompt composition, context reuse, cache-aware execution, and serving decisions. Workflow-aware Prompt Cache provides the current experimental evidence for this thread.

Workflow structurePrompt compositionKV-cache reuseContext and memoryModel servingRoutingInference optimization
Method

Transfer mechanisms, then test them under new assumptions.

  1. 01Classical systems mechanism
  2. 02Underlying assumptions
  3. 03Agentic workload characterization
  4. 04Broken assumptions
  5. 05Adapted or redesigned mechanism
  6. 06Implementation
  7. 07Benchmark and evaluation

We draw mechanisms from operating systems, networking, distributed systems, cloud computing, HPC, databases, storage, and resource management. Conceptual correspondence is not enough: a research contribution should lead to an implementable mechanism, measurable behavior, reproducible evidence, or a reusable artifact.

Shared Capabilities

Make results observable, comparable, and reproducible.

ObservabilityTracingWorkload characterizationBenchmarkingReproducibilityEvaluationCommon experiment infrastructure

HPC is an important mechanism source and experimental foundation for this work, not the boundary of our identity.

Evidence

Explore the systems and artifacts testing this research thesis.