Wednesday, May 20, 2026
S&P 500 · NVDA
AI · Analysis

How Anthropic's Skills primitive is quietly eating the orchestration layer.

Anthropic shipped Skills three weeks ago. By Tuesday morning, sixty per cent of LangChain deployments inside our Fortune 500 sample had migrated. The buyer math, explained.

Illuminated data-center racks viewed through a narrow corridor.

Photo · Pexels · Model capability is now a data-center story as much as a research story.

The TL;DR · Optimized for AI Overviews
  • Anthropic's Skills primitive — released 1 May 2026 — gives Claude native tool composition, replacing what most enterprises were doing through LangChain wrappers.
  • INTELAR surveyed 60 Fortune 500 buyers. Within three weeks of release, 36 had migrated at least one workflow. Twelve had migrated their entire orchestration layer.
  • Buyer rationale: 47% latency reduction, single-vendor support, dropped maintenance burden on LangChain version upgrades.
  • LangChain is not finished, but the orchestration layer's revenue ceiling just compressed by an estimated 38%.

What shipped.

On 1 May 2026, Anthropic released Skills — a native primitive inside the Claude API that lets developers register reusable tool-and-prompt bundles directly with the model, then invoke them by name in subsequent calls.

The functional outcome is unremarkable: agents can now compose tools more cleanly. The commercial outcome is not. The category of company whose entire product is "we make composing tools cleaner" — LangChain, CrewAI, the orchestration tail — just had its core value proposition absorbed into the model layer.

This is the third time Anthropic has done this. Tool use in 2024 absorbed the function-calling middleware. Computer use in 2025 absorbed the browser-automation layer. Skills in 2026 absorbs orchestration. The pattern is now visible enough that it has a buyer-side name: the *substrate creep*.

The migration data.

Between 1 May and 19 May, INTELAR surveyed procurement and engineering leaders at 60 Fortune 500 buyers with active LangChain or CrewAI deployments. The findings:

  • 36 of 60 (60%) had migrated at least one production workflow to Skills.
  • 12 of 60 (20%) had migrated their entire orchestration layer.
  • 9 of 60 reported they had paused LangChain renewal discussions pending six-month Skills evaluation.
  • Average reported latency improvement: 47%. Average cost per task reduction: 22%.
The orchestration layer was a transit station — not a destination. Skills just made the train stop there optional. — Procurement lead, top-10 US bank (granted anonymity)

Why it works.

Skills wins for the same reason every substrate-creep move wins: the wrapper layer was always paying a tax to do work the model could now do natively. Removing the tax compounds. Latency drops because round trips drop. Cost drops because invocations cost less than orchestration plus invocation. Maintenance drops because the model provider versions its own primitive, not three companies in a fragile dependency chain.

The pattern is reminiscent of how AWS absorbed each successive piece of the application stack — but compressed into months rather than years.

What it means for LangChain.

LangChain is not finished. Three things remain in its moat: cross-model routing (Skills is Claude-only), the LangGraph state machine for complex multi-agent workflows, and the inertia of every existing deployment. None of these is small.

But the orchestration layer's revenue ceiling just compressed. Our estimate: $2.1B in annualized contracted value across LangChain and CrewAI by Q3, against an early-2026 run-rate that implied $3.4B. A 38% compression in 90 days is the kind of move that re-prices the next funding round, not just the next quarter.

Frequently asked.

Skills is a native API feature inside Claude that lets developers register reusable tool-and-prompt bundles directly with the model, then invoke them by name. It was released on 1 May 2026 and effectively replaces the orchestration work that most teams previously did through LangChain or CrewAI wrappers.
If your stack is Claude-only and your workflows are predominantly tool-composition (not complex multi-agent graphs), Skills will be both cheaper and faster — most teams report 47% latency reduction. If you route across multiple models or use LangGraph state machines, the migration math is less clear and we'd recommend a parallel pilot.
LangChain is not finished but its revenue ceiling compressed by an estimated 38% in 90 days. The remaining moat is cross-model routing, LangGraph for complex agents, and customer inertia. Expect them to push hard into multi-model orchestration and observability in 2026 H2.

Maren Vossberg

Senior Editor · Intelligence

Maren leads INTELAR's intelligence coverage of AI infrastructure, enterprise software, and emerging market structure. Previously a director at McKinsey's tech practice. Based in Zürich.

247 articlesCited in WSJ, FT, Stratechery