The world is becoming agentic.

WHAT WE SEE

Agentic systems are becoming a new class of computational workload.

Their execution can emerge and adapt at runtime, changing assumptions that conventional systems were designed around.

01—

Execution emerges at runtime

Agents can choose tools, delegate work, create subagents, and form new execution paths while running.

02—

State persists across actions

Context, memory, tool state, intermediate artifacts, and external environments continue to shape what happens next.

03—

Resource demand changes with the workflow

Models, compute, memory, network, tools, and budgets depend on decisions made during execution.

04—

Infrastructure sees fragments, not the evolving workload

Schedulers, serving systems, and observability tools usually manage jobs, requests, models, or processes individually—not the agent workflow as a first-class workload.

OUR THESIS

Re-examine systems mechanisms for agent-native workloads.

01

Represent the workload

Develop abstractions for dynamic agent workflows, subagents, dependencies, state, and execution traces.

02

Transfer systems mechanisms

Start from operating systems, networking, distributed systems, HPC, databases, and resource management; identify which assumptions break.

03

Adapt or redesign

Turn the changed assumptions into concrete mechanisms for execution, scheduling, memory, inference, communication, and reliability.

04

Build and measure

Implement working systems and evaluate them with traces, baselines, benchmarks, reproducible experiments, and reusable artifacts.

RESEARCH AGENDA

Three connected systems research threads.

We study agentic execution models, runtime and resource management, and memory and inference systems as connected parts of the same workload shift.

HPC and scientific workflows provide a demanding mechanism source and experimental foundation, but they do not define the boundary of Agentivium AI.

EVIDENCE

What we have today.

Concrete systems and experiments that test parts of the broader research thesis.

INFRASTRUCTURE

Agentivium Core

Reusable contracts for intent, policy, memory, verification, execution traces, and evaluation.

RESEARCH INFRASTRUCTUREVIEW PROJECT →
SYSTEM

HPC-Claw

User-space HPC decision support with evidence, verification, approval, and guarded execution.

ACTIVE RESEARCHVIEW PROJECT →
EXPERIMENT

Workflow-aware Prompt Cache

Workflow-aware prompt composition and KV-cache reuse, with baselines and evaluation data.

RESEARCH PROTOTYPEVIEW PROJECT →
JOIN US

Build the systems beneath the agentic shift.

We welcome students and collaborators interested in agentic execution, runtime and resource management, memory and inference systems, and rigorous systems experimentation.