The big idea
Between late 2024 and mid-2026, the center of gravity moved three times.
**Chat → Agents.** The unit of value stopped being 'a good answer' and became 'a completed task.' AI went from giving advice to taking action.
**Prompts → Skills.** Anthropic formalized a portable, open-standard skill format (standardized Dec 2025) that now runs on Claude, OpenAI Codex, Cursor, Gemini CLI, Antigravity, and Windsurf. A skill is a reusable instruction set — think of it as a saved SOP your agent discovers and loads automatically.
**Integrations → MCP.** MCP (Model Context Protocol) became the universal plug format connecting AI agents to real tools. Over 1,000 production servers now exist, with official integrations from HubSpot, Salesforce, Clay, Apollo, Google, and Meta.
Those three primitives map directly onto the 'AI operating system' offer you sell: SOPs become skills, dashboards become MCP-powered reports, and quality control becomes eval loops with human review gates.
Numbers that matter
- 1,000+ Production MCP servers listed across Smithery, Glama, and PulseMCP directories. (MCP directory census, mid-2026)
- 5–20× Cost premium teams pay when they don't route across model tiers. (Field-manual synthesis §1.3)
- 3 Primitives that now map 1:1 to the 'AI OS' offer: skills, MCP, evals+hooks.
Deep dive
The three shifts, concretely
**Chat → Agents** happened when tools like Claude Code, Cowork, and browser-use agents let the model operate a computer instead of just answering questions. The model became a doer.
**Prompts → Skills** happened when Anthropic formalized the skill format: a folder containing a SKILL.md file (with a short YAML header and Markdown instructions), plus optional scripts and reference documents. *Progressive disclosure* is the key mechanic — roughly 100 tokens of name and description are always visible to the model; the full instructions load only when the skill is relevant; scripts load only on demand. This keeps context lean even when you have hundreds of skills installed.
**Integrations → MCP** happened when MCP went from an Anthropic side project (Nov 2024) to a cross-vendor standard with 1,000+ production servers by Q2 2026.
Five surfaces to be fluent in
**Claude.ai / Desktop** — where you do your thinking, drafting, and live client conversations.
**Claude Code** — the core agent harness. It reads and writes files directly, loads skills automatically, connects to MCP servers, and can run without you (headless mode). Don't let the 'Code' name mislead you — this is a general-purpose agent shell.
**Claude Cowork** — Claude Code with the terminal hidden. It's the surface you hand to clients so they get the power without the intimidating interface.
**Browser agents** — Chrome plus a developer browser for tasks that require logging in to a website (ad platforms, CRMs, etc.).
**The Claude API + Agent SDK** — for pipelines that run on a schedule or trigger, completely without human supervision.
Why model choice matters less than it used to
Think of models in three tiers, not as one monolithic decision:
**Frontier models** (each provider's current highest-capability model — names change fast, so check the linked model catalog) handle judgment-heavy work — strategy, orchestration, client-facing writing.
**Mid-tier models** (e.g. the current Sonnet-class model) handle the workhorse 80% — enrichment, drafting, subagent tasks.
**Small/fast models** (e.g. the current Haiku-class model, or local models via Ollama) handle high-volume, low-stakes calls — tagging, routing, data extraction.
The real skill is *routing*: assigning the right tier to each step in a workflow. Tools like n8n, Clay, and custom harnesses all support per-step model assignment. Teams that don't route pay 5–20× more per outcome with no quality gain.
The AI-nativeness ladder as diagnostic
**Level 0 — Tourist:** Uses ChatGPT to draft things, then manually copies the output. No real integration.
**Level 1 — User:** Has Claude Projects set up, uses prompt templates, saves good outputs as references.
**Level 2 — Builder:** Has built automations in n8n or Zapier, called APIs, maybe set up basic retrieval (RAG). This is where most 'AI-forward' agencies sit today.
**Level 3 — Operator:** Uses an agent harness as a daily driver. Has a skills library. Has MCP servers connected to the live stack. Has evals (automated quality checks) running on critical workflows.
**Level 4 — Architect:** Builds multi-agent systems, writes custom MCP servers, manages permission matrices and cost-routing policies, and deploys the whole setup as a client-ready operating system.
Levels 0–2 are fully commoditized. The commercial gap is at 3–4, and it's widening because each engagement adds reusable capability that compounds.
How to run it
- Name the harness. Pick one primary agent harness per engagement (Claude Code default) and document why.
- Map the surfaces. For each client task, choose the right of the five surfaces before writing a single prompt.
- Set routing policy. Assign a model tier (frontier/mid/small) per pipeline step and justify each in a comment.
- Score the ladder. Place the client on 0–4 and identify the one-rung-up artifact list.
In the field
Model routing in a lead pipeline: Haiku-class for initial filtering, Sonnet-class for enrichment synthesis, frontier only on the final 5% that merit a bespoke research brief. Teams that don't route this way pay 5–20× more per outcome for no quality gain.
When not to use it
If the client is still at Level 0–1 (tourist/user), don't sell a level-4 architecture. The ladder — Tourist, User, Builder, Operator, Architect — is a diagnostic; meet them one rung above where they are today, and price the compounding upgrades.
Pitfalls and takeaways
- Selling 'a custom GPT + some Zaps' — that's levels 0–2 and fully commoditized.
- Treating model choice as identity instead of a routing decision inside workflows.
- Ignoring that skills+MCP+harness engineering compounds — every engagement adds reusable capability.
- The unit of value is a completed task, not a good answer.
- Skills are portable across every major agent stack — no vendor lock-in.
- MCP is the integration layer; write access is the dividing line.
- Model choice is routing, not identity — assign per step.
- Levels 3–4 compound; levels 0–2 are commodity.
Do the work
Goal: Build a real AI-nativeness scorecard for one actual client (or your own agency) and turn it into a priced 90-day upgrade roadmap — not a hypothetical exercise, a document you could send today. (55 min)
- Pick the subject and gather evidence. Choose one client engagement (or your own shop) you know well enough to score honestly. List every AI-related artifact currently in use: prompt templates, automations, dashboards, agent harnesses, skills, MCP connections.
- Score each of the five dimensions against the ladder. For each ladder rung (Tourist, User, Builder, Operator, Architect), write one sentence of evidence for why the subject does or does not clear that rung. Land on a single 0-4 score with justification.
- Name the one-rung-up artifact list. For the gap between current and next rung, list the specific missing artifacts (e.g. "no skills library", "no MCP write access", "no eval harness on critical workflows", "no permission matrix"). Be concrete — name real tools and files, not categories.
- Price the climb. For each missing artifact, estimate build time and assign a sprint price anchored to replaced-headcount or time-saved value (not hours billed). Sequence them into a 90-day plan with 3 milestones.
- Write the one-page scorecard. Compress steps 2-4 into a one-page document: current rung + evidence, target rung, the artifact gap list, the 90-day milestone plan with prices. This is the actual sales/planning artifact.
- Pressure-test with a second opinion. Have Claude or ChatGPT role-play a skeptical client reading your scorecard and list every objection it would raise. Revise the document to pre-empt the two strongest ones.
Deliverable: A one-page AI-nativeness scorecard for a real subject: current rung with evidence, named artifact gaps, and a priced 90-day milestone roadmap that survived a skeptic pass.
Sources
- GMS AI-Native Field Manual 2026 §1 (The 2026 Landscape)
- Anthropic Agent Skills open-standard announcement (Dec 2025)
- Anthropic models overview (docs.claude.com)
- OpenAI models (platform.openai.com/docs/models)