Projects

Research systems for agent-native workloads.

Agentivium Core, HPC-Claw, and Workflow-aware Prompt Cache: reusable contracts, verifiable HPC decisions, and inference experiments.

Current Projects

Systems and experimental artifacts.

01Research infrastructure

Agentivium Core

Domain-neutral contracts

Research systems repeatedly need contracts for intent, evidence, policy, state, execution, and evaluation.

Agentivium Core provides typed schemas and reusable interfaces for intent parsing, policy validation, planning, tool abstraction, provider-neutral LLM calls, execution tracing, and evaluation. It also provides contracts for stateful assistants, memory, retrieval, and verification. Domain extensions supply scheduler logic, provider clients, policies, and orchestration; Core is not a complete agent runtime or orchestration platform.

Current work

Intent and policyPlanning and toolsProvider-neutral LLM callsStateful assistantsMemory and retrievalVerificationTracing and evaluation
02Open-source system / active research project

HPC-Claw

Verifiable HPC decision support

Agent-generated scheduling decisions need to be checked against scheduler state, policy, constraints, and evidence before they affect a cluster.

HPC-Claw is a user-space, pre-submission decision-support and verification layer. It handles structured or natural-language intent, captures Slurm context and policy snapshots, generates candidate decisions and evidence bundles, and checks verification and admissibility. Its CLI supports explicit approval, guarded execution, and execution audit, with dry-run as the default. HPC-Claw is the first domain extension of Agentivium Core and is not a scheduler replacement.

Current work

Slurm context and policyCandidate decisionsEvidence bundlesVerification and admissibilityApproval and guarded executionExecution audit
03Research prototype / experimental artifact

Workflow-aware Prompt Cache

Efficient multi-agent inference

Multi-agent workflows repeatedly reuse overlapping context. Can workflow structure improve cache reuse and reduce redundant inference?

This project investigates workflow-aware prompt composition and KV-cache reuse through an experimental implementation, multi-agent workflows, baselines, evaluation data, and analysis. It is a research prototype rather than a production cache system or serving platform.

Current work

Workflow structurePrompt compositionKV-cache reuseBaseline comparisonsEvaluation data and analysis
Contribute

Contribute to our research systems.