ontology agentic setup review for 2026-05-18

[github-actions]

Ontology Agentic Setup Review — 2026-05-18

Status snapshot across 5 watched ontology repositories. Evidence drawn from local dossiers generated on 2026-05-18.

Repo	Instruction files	Subagents	Skills	Agent workflows	Findings
EBISPOT/efo	1 (Copilot only)	3 (.github/agents)	0	1 (Copilot)	No CLAUDE.md; subagents unreachable by Claude
geneontology/go-ontology	2 (CLAUDE.md + Copilot)	0	8 (.claude/skills)	2 (Claude AI agent + Copilot)	Dual instruction files risk drift; no subagents
monarch-initiative/mondo	2 (CLAUDE.md + Copilot)	6 (.claude/agents)	4 (.claude/skills)	2 (Dragon AI agent + Copilot)	Most complete setup; dual file drift risk remains
obophenotype/cell-ontology	2 (CLAUDE.md + Copilot)	0	0	3 (Claude AI agent + CLARA + Copilot)	Most active in workflows but zero task decomposition
obophenotype/uberon	2 (CLAUDE.md + Copilot)	6 (.claude/agents)	0	3 (Claude AI agent + code review + Copilot)	Good subagent base; missing skills

Highest-leverage gaps: obophenotype/cell-ontology has 3 active agent workflows but no skills or subagents. EBISPOT/efo has 3 subagents that Claude cannot reach (Copilot-only instructions). monarch-initiative/mondo is the most complete model.

Generated by agent-watcher. Full comparative review in the comment below.

[github-actions]

Comparative Review — Ontology Agentic Setup (2026-05-18)

Central Agent Instructions

All repos except EBISPOT/efo have a CLAUDE.md at the repo root. EFO relies solely on .github/copilot-instructions.md, which means Claude-based agents (Claude Code, claude-code-action) run without project-specific context. EFO's .github/agents/ directory defines three subagents (EFO-curator.md, EFO-importer.md, EFO-ontologist.md) but these are scoped to Copilot and are inaccessible to Claude workflows.

Four repos (go-ontology, mondo, cell-ontology, uberon) maintain both CLAUDE.md and .github/copilot-instructions.md. In every observed case, the two files appear to carry identical content. This duplication creates a drift risk: edits applied to one file may not reach the other.

---

Task Decomposition (Skills and Subagents)

monarch-initiative/mondo is the most complete: 6 subagents in .claude/agents/ (deep-research-specialist, design-pattern-advisor, identifier-validator, metadata-checker, ontology-reasoner, and one more) plus 4 skills in .claude/skills/ (analyse-issue, merge-terms, odk, release-announcement). This covers both specialist delegation (subagents) and repeatable procedural tasks (skills).

geneontology/go-ontology has the richest skill library (8 entries: chemical-entity, design-pattern, external-term-lookup, mapping, reaction, and others) but no subagents at all. Skills are well-suited to GO's structured curation tasks, but complex multi-step operations would benefit from at least a few subagents.

obophenotype/uberon mirrors Mondo's subagent set (same 6 agents) but has no .claude/skills/ directory. The skill gap means Uberon's subagents cannot delegate to well-defined procedures for repeated tasks such as release announcements or ODK operations.

EBISPOT/efo has 3 subagents under .github/agents/ (Copilot-scoped) and no skills. Because the only instruction file is Copilot-specific, the agent decomposition is structurally sound for Copilot but invisible to Claude.

obophenotype/cell-ontology has neither subagents nor skills despite running three active agent workflows (AI agent, CLARA review, Copilot setup). This is the sharpest gap: agents are being triggered with nothing to delegate to.

---

Agent Workflow Coherence

Repo	Workflows	Agent targets
EBISPOT/efo	copilot-setup-steps.yml	Copilot only
geneontology/go-ontology	ai-agent.yml, copilot-setup-steps.yml	Claude Code + Dragon AI agent + Copilot
monarch-initiative/mondo	ai-agent.yml, copilot-setup-steps.yml	Dragon AI agent + Copilot
obophenotype/cell-ontology	ai-agent.yml, clara-review.yml, copilot-setup-steps.yml	Claude Code action + Copilot
obophenotype/uberon	ai-agent.yml, claude-code-review.yml, copilot-setup-steps.yml	Claude Code action + Copilot

GO ontology and Cell Ontology both use anthropics/claude-code-base-action or anthropics/claude-code-action for Claude-triggered automation. Uberon adds a dedicated claude-code-review.yml for PR code review. Mondo and EFO rely on Dragon AI agent (dragon-ai-agent/run-goose-obo) rather than Claude-native actions, meaning Mondo's rich .claude/ setup is currently targeting a different agent runtime.

---

Concrete Recommendations (by priority)

1. obophenotype/cell-ontology — add task decomposition immediately
Cell ontology has 3 active workflows triggering Claude agents but no .claude/agents/ or .claude/skills/ directory. At minimum, create a .claude/skills/ directory with skills for its most common curation operations (e.g., term addition, cross-reference lookup, PR review checklist). Uberon's subagent list is a ready-made template to adapt.

2. EBISPOT/efo — add a CLAUDE.md and wire subagents to Claude
EFO has well-defined subagents but they are Copilot-scoped. Add a CLAUDE.md (even a thin one pointing to the project guide) so Claude-based workflows have context. Migrate or link the three subagent definitions from .github/agents/ to .claude/agents/ to make them reachable by Claude Code action.

3. All dual-instruction repos — consolidate or auto-sync
For go-ontology, mondo, cell-ontology, and uberon: treat CLAUDE.md as the source of truth and either (a) auto-copy it to .github/copilot-instructions.md via a CI step, or (b) add a comment at the top of the Copilot file noting it is generated from CLAUDE.md. Without this, the two will diverge silently over time.

4. obophenotype/uberon — add skills to complement its subagent library
Uberon has 6 subagents (matching Mondo) but no skills. Adding Mondo's 4 skills (analyse-issue, merge-terms, odk, release-announcement) would give Uberon a nearly identical capability profile to Mondo, which is the current best-practice baseline.

5. geneontology/go-ontology — consider adding subagents
GO has the deepest skill library (8 skills) but no subagents. For complex cross-cutting tasks (e.g., taxon constraint validation, mapping review), named subagents would allow the main agent to delegate rather than handling everything inline. Mondo's deep-research-specialist and ontology-reasoner are natural starting points.

6. monarch-initiative/mondo — check Dragon AI agent vs Claude native alignment
Mondo uses Dragon AI agent (dragon-ai-agent/run-goose-obo) as its primary AI workflow, but the repo contains a rich .claude/ setup. Verify that the Dragon AI agent actually reads and benefits from the .claude/agents/ and .claude/skills/ directories; if not, consider adding a Claude-native ai-agent.yml alongside the Dragon workflow so the .claude/ investment is exercised.

---

Outlier Summary

• Most complete setup: monarch-initiative/mondo — subagents + skills + dual workflows
• Highest workflow activity, weakest decomposition: obophenotype/cell-ontology — 3 workflows, 0 skills/subagents
• Agent-platform mismatch: EBISPOT/efo — subagents defined for Copilot, no Claude context file
• Best skill library: geneontology/go-ontology — 8 skills, no subagents