Digital Talent Component Taxonomy
Every digital talent is assembled from 8 component categories. This taxonomy maps the building blocks, their selection criteria, relationships, and ownership.
The 8 Component Categories
Agentic Patterns
Reusable orchestration, reasoning, tool use, memory, and human-in-the-loop patterns that define HOW the talent works.
Frameworks
Domain methodologies and best practices that define WHAT the talent knows.
Models & Routing
Which LLM handles which task — cost, quality, privacy, and latency tradeoffs.
Harness Configuration
Claude Code runtime setup — CLAUDE.md structure, settings.json, hooks, permissions, memory.
Skills & Commands
Reusable skill library — the executable capabilities shipped with each talent.
Templates & Deliverables
Output formats, report structures, diagram templates. Driven by framework selection.
Integrations
MCP servers, external APIs, data sources, and tool connections.
Knowledge Base
Reference materials, content-in packages, domain knowledge, glossaries.
Component Relationships
Categories constrain and inform each other during assembly.
Harness Capability Matrix
Claude Code is both our factory tool and product deployment format. These are the configurable capabilities.
| Capability | What It Does | Maturity |
|---|---|---|
| CLAUDE.md | Agent identity, rules, workspace config | Mature |
| Skills | Domain-specific prompts with frontmatter | Mature |
| Hooks | PreToolUse / PostToolUse / Stop events | Mature |
| MCP Servers | External tool/API integration | Growing |
| Model Selection | Per-skill or per-talent routing | Growing |
| Subagents | Parallel/sequential task delegation | Growing |
| Memory | File-based persistence across sessions | Growing |
| Plugins | Shareable bundles of skills, hooks, agents | Gap |
| Agent SDK | Multi-agent orchestration (P-G-E) | Gap |
Harness Design Principles
From Anthropic's harness design article + factory experience.
1. Separate generation from evaluation
Agents that evaluate their own work produce mediocre output. Use separate evaluator agents with concrete grading criteria.
2. Context resets beat compaction
Structured handoffs via files provide a clean slate. Don't rely on conversation history for critical state.
3. Every component encodes an assumption
As models improve, some harness components become unnecessary. Stress-test assumptions at each major model release.
4. Complexity should match task complexity
Simple talents don't need orchestration. Complex talents benefit from planner-generator-evaluator architecture.
Maturity Roadmap
Now (Mature)
- CLAUDE.md structure
- Skill template standard
- Hook validation
- Framework packaging
- Pattern catalog (21)
Near-term (Fill Gaps)
- Model routing catalog
- Integration/MCP catalog
- Harness config reference
- Skill reuse across products
Next (R&D)
- Agent SDK multi-agent
- Planner-Generator-Evaluator
- Cross-model review
- Plugin-based delivery
Ownership & Governance
| Category | Primary Owner | Contributors |
|---|---|---|
| Agentic Patterns | Ada | Riley (new patterns from R&D) |
| Frameworks | Francois | Riley (new frameworks) |
| Models & Routing | Ada | Riley (model releases, benchmarks) |
| Harness Configuration | Ada | Ivan (infrastructure), Riley (new capabilities) |
| Skills & Commands | Pablo | Ada (pattern-skill alignment) |
| Templates & Deliverables | Francois | Pablo (production feedback) |
| Integrations | Ivan | Riley (new MCP servers) |
| Knowledge Base | Kai | Francois (framework KB), Fiona (client feedback) |
| Taxonomy (overall) | Ada | All above |