phData's six-level stack vs Anthropic's agent framing

phData-internal. Founder shared slide 8/22 of the AI Workforce Practice deck ("The Scoping Spine — the six-level stack") and asked whether it matches Anthropic's certification framing.

The slide

Persona → Use Case → Agentic Workflow → Agent → Skill → Tool. "Each level constrains the next. Skip one and you'll scope at the wrong altitude."

• Persona — the human role agents serve or embody
• Use Case — business problem with a measurable outcome
• Agentic Workflow — end-to-end sequence the agents perform
• Agent — discrete role with scoped responsibilities
• Skill — atomic capability the agent calls
• Tool — the function, API, or MCP the skill invokes

Accompanying narrative (founder's words): crawl (tools + instructions = first agent) → walk (more tools/skills) → run (orchestrator coordinating a workflow). Portfolio-of-agents approach at Lionsgate; build the upstream/downstream edges of winning agents and wire them together.

Where it matches Anthropic (verified vs claude-api reference, 2026-06-12)

• Tool = function/API/MCP and Skill = packaged on-demand capability are Anthropic's exact vocabulary (Agent Skills, MCP).
• Crawl-walk-run = Anthropic's official "start simple" posture: default to the simplest tier (single call → workflow → agent); only add agency for open-ended exploration.
• Anthropic's 4-question agent gate (complexity / value / viability / cost-of-error) is a sharper version of "tough to know which agent will take off."
• Anthropic's Managed Agents multi-agent API enforces one level of delegation (coordinator → workers, depth >1 ignored) — validates a single orchestrator layer, argues against orchestrator stacking. "Run" should be earned by observed friction, not reached as a stage.

The divergence (cert-relevant)

"Workflow" altitude collision. The slide puts Agentic Workflow ABOVE agents. Anthropic's canonical distinction: workflow = predefined, code-orchestrated control flow; agent = model directs its own loop. Siblings/alternatives, not layers — and workflow is usually the simpler tier below an agent. Cert (Claude Certified Architect – Foundations; Agentic Architecture = 27%, heaviest domain) will test the Anthropic sense. phData vocabulary in the deck, Anthropic vocabulary in the exam room.

The gap (client-pitch upgrade)

No layer in the stack owns measurement. "Measurable outcome" is named at level 2 but nothing measures. Anthropic productized this: Outcomes — user.define_outcome with a gradeable rubric; a separate grader model runs an iterate→grade→revise loop (max_iterations, per-criterion feedback). Upgrades the Lionsgate portfolio answer: which-agent-takes-off becomes a scoreboard, and the scoreboard picks which winner's edges to build next. Maps to the founder's four-layer targeting filter: slide top = targeting, slide bottom = tools; instrumentation + feedback loop are the missing middle ([[feedback_targeting_system_prioritization_filter]]).

Cross-links

• [[2026-06-09-caf-restructure-proposal]] — move 4 (collapse 106 prescriptive skills → objective+acceptance-test skills) is the same insight: acceptance tests are rubrics; CAF's missing layer is also measurement.
• [[2026-06-11-sales-process-dsa-training]] — the sales-process site (phdata-sales.rdco.dev) is the companion artifact.
• Lived validation of edges-of-winners: quick-sites (6/10) → home HQ (6/10) → phdata-sales site (6/11).