GitHubJoin Us
Research

Research in Agentic Computing.

Agentivium AI studies Agentic Computing as an emerging field: the computational models, abstractions, systems, and mechanisms underlying autonomous agent workloads. Agent-native systems are our design principle; systems and infrastructure are our technical focus.

Our central question is what changes when control flow is model-mediated and dynamically generated during execution. We investigate those changes through working systems, measurement, and reproducible experiments—not by treating a conceptual mapping as a result.

Research Agenda

Three connected threads.

01

Agentic Execution Models

How dynamically formed agentic execution should be represented and controlled: workflows, subagent graphs, execution semantics, spawning, communication, and state.

Focus

Dynamic execution graphsSub-agent spawningWorkflow representationDistributed agent runtimeRuntime orchestration

Research questions

  • How should dynamically formed agentic execution be represented and controlled?
  • Which execution semantics support spawning, communication, state capture, and replay?
  • Which operating-system and distributed-systems assumptions hold for agentic workloads?
02

Agentic Runtime & Resource Management

How runtime decisions can allocate models, compute, memory, tools, and budgets as agent workloads evolve under uncertainty.

Focus

Runtime predictionSchedulingMORLMulti-objective optimizationValue-per-computeAdaptive execution

Research questions

  • How can runtime behavior of agent workflows be predicted?
  • How should resources be allocated across quality, latency, cost, energy, and reliability objectives?
  • How should orchestration adapt online as workload structure and demand change?
03

Agentic Memory & Inference Systems

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

Focus

KV-cache reusePrompt compositionModel servingCache-aware executionRoutingInference optimization

Research questions

  • How can context and KV cache be reused across related workflows and agents?
  • How should serving and routing exploit workflow structure?
  • How do agent workflows change conventional inference-system assumptions?
How We Research

Transfer mechanisms, then test them under the new assumptions.

  1. 01Classical systems mechanism
  2. 02Changed workload assumption
  3. 03Adapted mechanism
  4. 04Implementation
  5. 05Benchmark and evaluation

We draw mechanisms from operating systems, networking, distributed systems, cloud computing, HPC, databases, and resource management. HPC is a mechanism source and experimental foundation, not the boundary of our identity.

Shared Capabilities

Make results observable, comparable, and reproducible.

ObservabilityTracingWorkload characterizationBenchmarkingReproducibilityEvaluationCommon experiment infrastructure

Supporting research assets, including Agentivium Core and Digital-X, can inform this shared infrastructure. They are not presented as separate flagship research threads or released product claims.

Explore Further

Explore the systems and artifacts behind our research.

View ProjectsJoin Us