Executable governance for agentic AI in research data lifecycle management

Knowledge graphs, propose→review→accept protocol, eight Tier-A validators, and a human review dashboard — not prompt text alone.

DOI 10.5281/zenodo.20650518 v3.0.0 MIT (code) CC BY-NC 4.0 (docs)
GitHub repository Zenodo archive Reproduce in 5 min ↓

Open-source research software by Boyuan (Keven) Guan (@Keven1894). License: MIT (code) · CC BY-NC 4.0 (docs & KG content).

Production context

Developed and battle-tested through production research data lifecycle management at an academic library: 10+ governed agents, 100+ TB under management, data types spanning historical aerial photos, GIS layers, time series, and real-time sensor streams; pipelines covering FAIR publication, data dashboards, GIS cataloging, real-time streaming, and MLOps across 14 active projects.

Patterns were generalized from EnviStor practice. Public repositories are framework-only — no production data, no Envita deployment code.

Reproduce in 5 minutes

Pin release v3.0.0 for paper reproducibility. Requires Python 3.11+.

git clone https://github.com/Keven1894/AgentLoom.git
cd AgentLoom && git checkout v3.0.0
python -m venv .venv

# macOS / Linux
source .venv/bin/activate
pip install -e .[dev]
python -m agentloom.kg.validate_all
python -m agentloom.validators.run_all

# Windows (PowerShell)
# .venv\Scripts\pip install -e .[dev]
# .venv\Scripts\python -m agentloom.kg.validate_all
# .venv\Scripts\python -m agentloom.validators.run_all
Expected output: all schema checks PASS for 6 KG files; PASS — all 8 Tier-A validator(s) succeeded.
Workshop quickstart workflow: clone, install, validate, open dashboard

Optional: start the review dashboard — python -m uvicorn agentloom.dashboard.app:app --port 8000 --host 127.0.0.1 → open http://127.0.0.1:8000. Full guide: docs/guides/START_HERE.md.

Ecosystem

Four repositories, four roles — split by what they are and by agent lifecycle phase.

AgentLoom ecosystem: co-agenticOS sets rules, AgentLoom is authoring-time framework, agentloom-runtime is run time, workshop repo is a concrete instance
co-agenticOS sets the rules → AgentLoom is the framework you build and govern an agent with → AgentLoom Runtime is the library the deployed agent uses → the workshop repo is one concrete instance.
Repository Role Lifecycle phase
co-agenticOS Governance spec — rules, coordination, memory boundaries sets the rules
AgentLoom Build framework — KG governance, validators, propose-review-accept authoring time
agentloom-runtime Runtime library — layered memory, graph-first retrieval run time
ucgis-agentloom-2026-workshop Concrete instance — worked, forkable example a use of all of the above

v3 architecture

Runnable stack: governed KG data, executable validators, human review dashboard, CLI accept gate.

AgentLoom v3 architecture: knowledge graphs, framework services, human reviewer, Tier-A validators

Core workflow

Governance flow: agent proposes, human reviews on dashboard, human accepts via CLI, validators enforce

Protocol details: propose-review-protocol.md · Architecture: agentloom-architecture.md

Dual-helix architecture

Learning and governance strands interlock — engineering auto-extends; governance human-reviews.

Dual-helix architecture: Learning helix and Governance helix interlock to produce reliable agents

Full narrative: dual-helix-clarification.md · Diagram spec: dual-helix-diagram-spec.md · Governance spec: co-agenticOS

Relationship to EnviStor

AgentLoom is a domain-neutral governance framework. EnviStor and Envita are institutional instances that apply these patterns in production research data management — they are not shipped in the public AgentLoom repository.

The agent governance layer in EnviStor-related work follows the AgentLoom framework. Framework source is public and domain-neutral; EnviStor/Envita deployment artifacts and production data are not included.

For an institutional case study (DIVA / Dataverse), see the Dataverse DIVA case study in the repository docs.

Research & citation

Software availability (papers)

AgentLoom v3.0.0 is archived at https://doi.org/10.5281/zenodo.20650518.
Reproduction steps and ecosystem overview: https://keven1894.github.io/AgentLoom/

BibTeX

@software{guan2026agentloom,
  author       = {Guan, Boyuan (Keven) and Juh{\'a}sz, Levente and Cui, Wencong},
  title        = {AgentLoom: Executable Governance Framework for Builder Agents},
  year         = {2026},
  version      = {3.0.0},
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.20650518},
  url          = {https://github.com/Keven1894/AgentLoom}
}

Version-pinned cite: 10.5281/zenodo.20650518 (v3.0.0). Concept DOI (latest): 10.5281/zenodo.17561540. Author ORCID: 0000-0002-4244-5011. See also CITATION.cff.

Related research

Contributors

Dr. Boyuan (Keven) Guan

Framework creator & lead developer

@Keven1894 · ORCID

Dr. Levente Juhász

UCGIS 2026 workshop co-organizer · University of Florida, GATOR Lab

@jlevente

Dr. Wencong Cui

UCGIS 2026 workshop co-organizer

@Vincent-Cui

Full list: CONTRIBUTORS.md