Skip to content

5. Agentic AI Skills for DOME

What happens when you give a frontier model the tools, not just the prompt.

Overview

Questions:

  • What does an “agent skill” add over a single LLM call like DOME Copilot?
  • What does the DOME Agent Skill actually do, phase by phase?
  • How does it produce DOME and OSAI output in one pass?
  • What does the extra capability cost, and who can realistically access it?

Learning Objectives

By the end of this chapter, you will be able to:

  1. Explain the difference between a single-shot LLM annotation and an agentic pipeline
  2. Walk through the eight phases of the DOME Agent Skill
  3. Weigh the accuracy, cost and accessibility trade-offs against DOME Copilot

Time: 10 minutes

Prerequisite: Chapter 3 — DOME Copilot and Chapter 4 — OSAI

From a model call to an agent

DOME Copilot does one thing very efficiently: it takes the text you give it and returns a structured annotation. The agentic approach changes the shape of the problem — instead of one model call over one document, a frontier model is given tools and left to chain them:

DOI or URLwhat you supply
Resolve metadataCrossRef, Europe PMC, arXiv, bioRxiv, Zenodo
Fetch full text+ supplementary files
Extract textPyMuPDF, page-marked
Fill 21 DOME fieldseach tagged with its source
Inspect linked repolightweight external check
Suggest OSAI componentsone fix per gap
Human reviewrequired — you correct every field
DOME Registrysubmission not enabled yet

Three things follow from that shape:

  1. Ground-truth metadata comes from APIs, not the model. Title, authors, journal, DOI, PMID are resolved through CrossRef, Zenodo, arXiv, bioRxiv and Europe PMC — so they are looked up, not recalled, and cannot be hallucinated.
  2. Supplementary material is in scope. Many DOME fields — data splits, hyperparameters, evaluation files — live in supplementary PDFs rather than the main text. An agent can fetch and read them.
  3. DOME and OSAI in one pass. Once the gaps are known, the agent queries the OSAI ecosystem list and returns concrete components that would close each one. Copilot does not do this.

A frontier model’s large context window is what makes this practical: the whole paper, its supplementary files and the schema can be held at once, which reduces the number of amendments a human has to make afterwards.

The DOME Agent Skill

gavinf97/dome-agent-skill is an open implementation of exactly this — v0.1.0, CC BY 4.0, public 1. The deterministic API and file work is done by six small Python CLIs; the DOME field extraction is done by the agent reading the paper.

The eight phases

# Phase What happens
1 Input Paper URL/DOI, any code repository links, and which schema version to target (default v2.0.0)
2 Resolve metadata resolve_publication.py — CrossRef → Zenodo → arXiv → bioRxiv/medRxiv → Europe PMC, cross-referenced via the NCBI ID Converter
3 Fetch full text fetch_fulltext.py — Europe PMC full-text XML plus supplementary files, filtered to documents under 20 MB
4 Extract text extract_pdf_text.py — PyMuPDF, with page markers so findings stay citable
5 Fill 21 DOME fields The agent reads the text and fills the schema, guided by a per-field extraction guide
6 External check inspect_external_link.py — one API call plus README for a linked repo. Deliberately lightweight: no cloning, no walking the file tree
7 Notes & compliance summary Provenance breakdown, remaining gaps, and OSAI-grounded improvement suggestions via suggest_osai_components.py
8 Human review The full entry and summary are presented for correction. Submission to the Registry is not enabled

What makes the output trustworthy

Source provenance on every field. Each filled field is tagged with where the evidence came from:

  • paper — stated in the manuscript itself
  • supplementary:<filename> — found in supplementary material
  • external:<url> — found only in a linked repository

That last tag matters more than it looks. DOME assesses what the paper discloses, not what a linked repo happens to contain. A field that could only be filled from GitHub is explicitly called out in the summary as “not disclosed in the paper itself” — which is a DOME finding, not a DOME pass.

The trade-off, stated plainly

Neither route is simply better. They optimise for different things.

DOME Copilot DOME Agent Skill
Engine Mistral Small 3.1 24B, self-hostable Frontier model (Claude or comparable)
Scope Main text → DOME fields DOI → metadata, full text, supplementary, linked repo → DOME + OSAI
Speed ~2 min ~5–15 min
Cost per paper Very low; compute-efficient Higher — you pay for frontier model tokens
Accessibility Web UI, nothing to install Needs Python 3.10+, a terminal and an agent platform
Scales to a corpus Yes — this is its purpose Not economically, at present
Human amendments needed More Fewer, in practice
Published benchmark Yes 2 No

Sources and further reading

Next: Hands-on: Submission Routes

  1. Gavin Farrell. DOME agent skill. 2026. Licensed CC BY 4.0. URL: https://github.com/gavinf97/dome-agent-skill. ↩↩

  2. Gavin Farrell, Omar Abdelghani Attafi, Styliani-Christina Fragkouli, Ignacio Heredia, Saul Fernández Tobías, Melissa Harrison, Henning Hermjakob, Matt Jeffryes, Marta Obregón Ruiz, Matt Pearce, Nikos Pechlivanis, Alvaro López García, Fotis Psomopoulos, and Silvio C. E. Tosatto. DOME copilot: making transparency and reproducibility for artificial intelligence methods simple. bioRxiv, 2026. Preprint, posted 19 April 2026. doi:10.64898/2026.04.16.718888. ↩