Date: 2026-04-11 Scope: Complete audit of every skill touchpoint across the Oshun monorepo, cross-referenced with April 2026 SOTA in AI skill systems, auto-generation, composition, evaluation, and improvement.
Executive Summary#
The Oshun monorepo contains 30+ independent skill-like systems across its 45+ domains — from Iris agent tools to Maya game abilities to Nous training methods to Veritas newsroom agents. However, these systems are entirely siloed with no unified framework, no cross-domain composition, no automatic skill generation, no autonomous skill library growth, and no skill-specific quality tracking.
The AI skills landscape has undergone a revolution in 2025-2026:
- Agent Skills Open Standard (Anthropic, Dec 2025): SKILL.md format adopted by 16+ major AI tools
- Memento-Skills (March 2026): Agent-designing agents that autonomously grow skill libraries (235 skills, doubled HLE performance)
- OpenSpace (March 2026): Self-evolving skill engine via MCP, 4.2x performance improvement
- MCP-Zero (June 2025): Active tool discovery via RL, 98% token reduction
- SkillX (April 2026): Three-tiered hierarchical skill knowledge (planning/functional/atomic)
- Skills.sh (Vercel, Jan 2026): Skill marketplace with 147 new skills/day
- xLAM-2 (Salesforce): SOTA open-source tool-use models, #1 on BFCL
Oshun has none of these. The gap between current state and SOTA is massive.
1. Current Skill Systems Inventory#
1.1 Where Skills Exist Today (30+ Systems Found)#
| Domain | Skill System | Skills Count | Has Registry | Has Versioning | Has Quality Tracking | Has Auto-Generation |
|---|---|---|---|---|---|---|
| Iris | Agent tool framework | 50+ tools | Yes | Yes | Partial | No |
| Iris | Computer-use agent | 20+ actions | No | No | No | No |
| Iris | Agent archetypes | 5 archetypes | No | No | No | No |
| Psyche | Avatar expression skills | 50+ expressions | No | No | Partial | No |
| Psyche | Action safety framework | 20+ policies | No | No | Yes | No |
| Psyche | Accessibility skills | 15+ checks | No | No | Yes | No |
| Hathor | NPC behavior/dialogue | 30+ behaviors | No | No | Partial | No |
| Hathor | Quest/narrative skills | 20+ operations | No | No | Partial | No |
| Maya | Character abilities | 100+ abilities | Yes (tree) | Yes (ranks) | Yes (balance) | No |
| Maya | Skill tree system | 50+ nodes | Yes (graph) | Yes (ranks) | Yes | No |
| Maya | Combat skills | 30+ mechanics | No | No | Yes | No |
| Maya | Crafting recipes | Hundreds | Yes | No | Yes (success rate) | No |
| Maya | NPC Souls behaviors | 76 Rust modules | No | No | Partial | No |
| Nous | Training methods | 50+ methods | No | No | Yes (loss) | No |
| Nous | Safety checks | 25+ detectors | No | No | Yes | No |
| Isis | Generation workflows | 20+ workflows | Yes | Yes | Yes | No |
| Isis | AI provider skills | 15+ providers | Yes | No | Partial | No |
| Veritas | Newsroom agent roles | 11+ agents | No | No | Partial | No |
| Veritas | Content generation | 10+ operations | No | No | Partial | No |
| Metis | Educational agents | 10+ agents | No | No | Partial | No |
| Metis | Assessment skills | 10+ types | No | No | Yes | No |
| Yemaya | Production pipeline steps | 30+ steps | Yes | No | Yes | No |
| Yemaya | QA validator skills | 50+ validators | No | No | Yes | No |
| Maat | Strategic analysis skills | 20+ analyses | No | No | Partial | No |
| Bellona | Code generation skills | 40+ generators | No | No | Partial | No |
| Euterpe | Music generation skills | 20+ operations | No | No | Partial | No |
| Shakti | Fitness training skills | 30+ exercises | No | No | Yes | No |
| Aglaea | Beauty technique skills | 20+ techniques | No | No | Partial | No |
| Aphrodite | Game strategy skills | 15+ strategies | No | No | Yes | No |
| Uzume | Live production skills | 50+ operations | No | No | Partial | No |
Key Finding: Every domain has skills but NONE has: auto-generation, cross- domain composition, autonomous growth, or standardized quality tracking.
1.2 Existing Skill Infrastructure#
Iris is the most mature with:
ToolId,ToolVersion,ToolSchema,ToolDefinitiontypesIToolDiscoveryinterface with search(), recommend(), discoverFrom(), scan()ToolCategorytaxonomy (filesystem, web, code, api, database, etc.)ToolCapabilityflags (stateless, stateful, async, streaming, etc.)- Composable pipelines with
ToolStepNode,ParallelToolNode,ConditionalToolNode ToolLimitsandToolSecurityRequirements
All other domains have ad-hoc skill structures with no shared framework.
2. Gap Analysis#
GAP 1: No Unified Skill Framework (CRITICAL)#
30+ independent skill systems with incompatible interfaces. No way to:
- Discover skills across domains
- Compose skills from different domains
- Track quality uniformly
- Share skills between domains
GAP 2: No Autonomous Skill Library Growth (CRITICAL)#
SOTA systems (VOYAGER, Memento-Skills, OpenSpace) autonomously discover, create, verify, and grow skill libraries. Oshun has zero autonomous skill generation.
GAP 3: No Skill Auto-Generation from Documentation (HIGH)#
Gorilla/ToolFactory/APIGen can generate skills from API docs and OpenAPI specs. Oshun has dozens of internal APIs but no doc-to-skill pipeline.
GAP 4: No MCP Skill Integration (HIGH)#
MCP is the de facto protocol (5,800+ servers, donated to Agentic AI Foundation). Oshun has partial MCP integration in Iris but no MCP-based skill ecosystem.
GAP 5: No Skill Quality Tracking Across Domains (HIGH)#
Each domain tracks quality differently (Maya has balance metrics, Nous has loss curves, Iris has execution monitoring). No unified quality framework.
GAP 6: No Skill-Specific Fine-Tuning (HIGH)#
xLAM-2 (Salesforce) shows SOTA tool-use model training. Oshun has no pipeline for fine-tuning models on domain-specific skill execution data.
GAP 7: No Skill Composition Across Domains (HIGH)#
Cannot compose "research (Sophia) → generate image (Isis) → verify quality (Nous) → publish (Yemaya)" as a single composed skill.
GAP 8: No Skill Marketplace/Registry (MEDIUM)#
Skills.sh has 147 new skills/day. ClawHub has 3,286 versioned skills. AWS Agent Registry launched April 2026. Oshun has no equivalent.
GAP 9: No Process Reward Models for Skill Execution (MEDIUM)#
AgentPRM/ToolPRMBench show step-level rewards for multi-step tool use produce better agents than outcome-only rewards. Not implemented.
GAP 10: No Self-Evolving Skill System (MEDIUM)#
OpenSpace's three evolution modes (FIX, DERIVED, CAPTURED) and Memento-Skills' agent-designing-agent pattern have no equivalent.
3. Domain-by-Domain Skill Gaps#
3.1 Iris (Conversational AI)#
- Has: Tool registry, discovery, composition, pipelines
- Missing: VOYAGER-style autonomous tool creation, MCP skill wrapping, tool-use fine-tuning data generation, process reward models for tool chains, skill versioning lifecycle, tool quality regression testing, self-improving tool selection
3.2 Psyche (Virtual Assistant)#
- Has: Expression skills, accessibility checks, action safety
- Missing: Skill transfer between avatar/voice/perception subsystems, computer- use skill auto-generation from UI analysis, meeting skill auto-discovery
3.3 Maya (Game Engine)#
- Has: Richest skill system (abilities, skill trees, combat, crafting, NPC Souls)
- Missing: Dynamic ability rebalancing from player data, ML-based combo discovery, NPC skill learning from player behavior, procedural ability generation
3.4 Nous (Training Platform)#
- Has: 50+ training methods, 25+ safety checks
- Missing: Auto-selection of best training method per task, training recipe composition, safety check ensemble optimization
3.5 Isis (Generative Factory)#
- Has: Workflow registry, provider skills, quality pipelines
- Missing: Auto-workflow generation from task description, learned parameter optimization, workflow quality prediction
3.6 Veritas (News Agency)#
- Has: 11+ newsroom agents with role-based skills
- Missing: Agent skill improvement from article performance, editorial skill transfer, beat-specific skill specialization
3.7 All Other Domains#
- Mostly ad-hoc skill structures with no standardized definition, discovery, composition, quality tracking, or improvement mechanisms
4. Recommended New Phase: Skills System (Phase 97)#
4.1 Unified Skill Framework#
Build a universal skill definition, registry, and runtime that all 45+ domains can adopt:
- Canonical
UniversalSkillschema (name, parameters, handler, quality metrics, dependencies, versioning, security) - Skill registry service with semantic search and discovery
- Skill composition engine (sequential, parallel, conditional, loop patterns)
- Skill quality tracking (success rate, latency, cost, user satisfaction)
- Skill versioning with semver, rollback, and migration
- Cross-domain skill composition (Sophia research → Isis generation → Nous safety)
4.2 Autonomous Skill Generation & Growth#
Implement VOYAGER/Memento-Skills/OpenSpace patterns:
- Automatic skill generation from API documentation (Gorilla/ToolFactory pattern)
- VOYAGER-style autonomous curriculum for skill discovery
- Memento-Skills pattern for agent-designing-agent (skills stored as markdown, organic growth from execution experience)
- OpenSpace-style self-evolving skill engine (FIX, DERIVED, CAPTURED modes)
- MCP-Zero active tool discovery (RL-trained tool request generation)
- CREATOR-style LLM tool creation (generate both tool code and documentation)
4.3 Skill-Specific Training & Evaluation#
Build training and evaluation infrastructure for skills:
- Skill execution data collection for fine-tuning (data flywheel connection)
- xLAM-style tool-use model training pipeline
- APIGen-MT synthetic multi-turn tool-use data generation
- Process reward models for multi-step skill execution (AgentPRM/ToolPRMBench)
- BFCL-style skill evaluation benchmark
- Skill degradation detection and auto-improvement
4.4 MCP Skill Ecosystem#
Build the MCP-based skill infrastructure:
- MCP server for every Oshun domain (expose domain capabilities as MCP tools)
- MCP skill wrapping (MCP tool + SKILL.md prompt template = composable skill)
- OpenAPI-to-MCP auto-generation for all internal APIs
- MCP skill marketplace (internal registry + external community skills)
- MCP skill composition across domains
4.5 Per-Domain Skill Enhancement#
Enhance each domain's skill system with the unified framework:
- Iris: VOYAGER-style tool library growth + MCP integration + tool-use fine-tuning
- Maya: Dynamic ability generation + NPC skill learning + combo discovery
- Nous: Training recipe auto-selection + safety ensemble optimization
- Isis: Auto-workflow generation + parameter optimization
- Veritas: Editorial skill improvement from performance data
- All domains: Adopt unified skill definition + quality tracking
5. SOTA Systems to Integrate#
| System | What It Does | Priority |
|---|---|---|
| Agent Skills Standard (SKILL.md) | Universal skill format | P0 |
| Memento-Skills | Agent-designing-agent, organic skill growth | P0 |
| OpenSpace | Self-evolving skill engine via MCP | P0 |
| MCP-Zero | Active tool discovery via RL | P1 |
| SkillX | Three-tiered hierarchical skill knowledge | P1 |
| VOYAGER | Autonomous curriculum + skill library | P1 |
| xLAM-2 | SOTA tool-use model training | P1 |
| APIGen-MT | Multi-turn tool-use data synthesis | P1 |
| ToolFactory | Auto-generate tools from API docs | P1 |
| Gorilla/RAFT | Retriever-aware API calling | P2 |
| AgentPRM | Process reward models for tool agents | P2 |
| CREATOR | LLM creates its own tools | P2 |
| ReAcTree | Hierarchical agent trees for composition | P2 |
| OpenAI o3/o4 pattern | RL-trained native tool reasoning | P2 |
| Plan-and-Act | Separated planning/execution | P2 |
| Skills.sh integration | External skill marketplace | P3 |
| ClawHub integration | External skill registry | P3 |
| BFCL v4 | Function calling benchmark | P1 |
| tau-2-Bench | Agent-user interaction evaluation | P1 |
| MCPAgentBench | MCP tool use benchmark | P1 |