- Monday Momentum
- Posts
- From Chat to Toolchain
From Chat to Toolchain
Why LLM providers are building their own Integration Layer, and how Claude's Skills changes the game
Justin Wright
October 20, 2025 • Est. Reading Time: 9 minutes
Happy Monday!
Last week, I explored how OpenAI's AgentKit transforms workflow automation into agentic orchestration, serving different needs than established platforms like Zapier and n8n. But just days after AgentKit's launch, the competitive dynamics shifted again: Anthropic released Skills for Claude, a framework that packages instructions, scripts, and executable code into folders Claude automatically loads when relevant.
The pattern is unmistakable. Within a week, both major LLM providers launched tools that move beyond chat interfaces into complete application platforms. OpenAI built visual orchestration with AgentKit. Anthropic built progressive disclosure with Skills. Different architectures, same strategic goal: keep users inside their ecosystem for complete workflows instead of simply just conversations.
This isn't about replacing Zapier or building "better chat." It's about reimagining the software stack when AI is the primary interface. Skills let Claude run executable code for deterministic tasks, automatically compose multiple capabilities, and work across Claude apps, Claude Code, and the Claude API. The transformation: from answering questions to actually doing work.
Anthropic's Skills and OpenAI's AgentKit signal LLM providers building complete application stacks instead of just models. Skills introduce "progressive disclosure" where Claude selectively loads only relevant capabilities. This new software approach is optimized for agentic systems. Both platforms include executable code environments, transforming chat interfaces into complete workspaces.
Progressive Disclosure: A New Software Primitive
Anthropic introduced a concept that deserves more attention than it received: progressive disclosure as the "core design principle" for Skills. The architecture fundamentally differs from traditional software approaches, and may represent a new paradigm optimized for agentic systems.
Skills are folders containing instructions, scripts, and resources that Claude loads dynamically. But the innovation isn't just in the packaging, it's the selective loading mechanism. As Anthropic explains, "like a well-organized manual that starts with a table of contents, then specific chapters, and finally a detailed appendix, Skills let Claude load information only as needed."
The technical implementation reveals why this matters. Claude initially sees only skill names and brief descriptions, then autonomously decides which Skills to load based on the task, accessing only specific files and information needed at that moment. Each Skill contains a SKILL.md file with metadata that helps Claude determine when to activate it.
How This Differs from Existing Approaches:
Traditional software loads everything upfront; applications pull all dependencies, libraries, and configurations at startup. This creates more predictability but wastes resources on unused functionality.
RAG (Retrieval Augmented Generation) searches knowledge bases and retrieves relevant chunks when needed. But RAG is reactive; it responds to queries with information, not capabilities.
Skills introduce agent-driven selective loading where Claude can "contain an unbounded amount of context to teach Claude how to complete a task" because the agent intelligently navigates what to load. Skills can reference additional files, creating layered skill structures Claude reads only when relevant.
The efficiency gains matter for enterprise deployment. Skills keep Claude fast while accessing specialized expertise because irrelevant capabilities never load. This addresses the context window constraints that plague simple RAG systems while avoiding the bloat of monolithic applications.
The architectural implications extend beyond Anthropic. Progressive disclosure represents a new pattern for building agentic systems, one where capabilities remain dormant until contextually activated. Just as databases introduced indexes (selective data access) and caching introduced memoization (selective computation), Skills introduce context-aware selective loading as a tool for AI systems.
The Traditional Software Stack:
Frontend: HTML/CSS/JavaScript, user interfaces, forms
Backend: Business logic, APIs, server-side processing
Database: Persistent storage, queries, data management
Integration Layer: Zapier, n8n, middleware connecting systems
The Emerging LLM Stack:
Chat Interface: Natural language becomes the interaction layer
Model: Reasoning and decision-making happen in the LLM
Context Window: Conversation history and loaded Skills provide state
Skills/AgentKit: Native capabilities for connecting to tools and executing tasks
Anthropic emphasizes that Skills work across Claude apps, Claude Code, and the API, creating a unified execution environment. The company positions Skills as "create once, use everywhere", enabling the same capability to function whether accessed through chat, developer tools, or programmatic API.
Pattern Recognition: Code Execution as the Hidden Moat
While progressive disclosure provides architectural elegance, the strategic moat may lie in something more fundamental: Skills' ability to include executable code.
Anthropic explicitly acknowledges AI's limitations: "Skills can include executable code for tasks where traditional programming is more reliable than token generation." The company cites sorting lists as an example: a coded sorting algorithm completes the task faster, at less cost, and produces identical output every time compared to LLM token generation.
This hybrid approach, combining AI reasoning with deterministic code execution, addresses the core limitation that plagued our "workslop" discussion: AI generating output that looks right but lacks substance. Skills provide structured, reliable execution for tasks requiring precision while reserving AI for tasks requiring judgment.
For example, a PDF skill extends Claude's document understanding by adding form-filling or field extraction via scripts. The AI determines when the task requires these capabilities, loads the Skill, and executes the code, all without user intervention.
Skills acknowledge that well-designed code outperforms AI generation for deterministic tasks. The key becomes knowing when to use AI reasoning versus when to use traditional programming; architectural thinking that becomes more valuable, not less, in an AI-dominated world.
Contrarian Take: Integration Becomes Invisible, Not Irrelevant
The narrative that AgentKit and Skills "kill" traditional integration platforms misses a more subtle transformation: integration as a discrete category is disappearing, absorbed into the model layer itself.
Traditional integration platforms made explicit what was implicit. Zapier workflows visualized data flow: when this trigger fires in App A, execute this action in App B. The value came from making integration logic visible, manageable, and maintainable.
Skills and AgentKit represent the opposite philosophy: integration should be implicit, invoked contextually by intelligent agents rather than explicitly configured by humans. Claude scans available Skills to find relevant matches, loading capabilities automatically. AgentKit coordinates multi-agent workflows without users manually mapping connections.
This doesn't eliminate the need for integration, it just changes who (or what) manages it. The agent becomes the orchestrator, deciding which tools to invoke and when to invoke them. Integration logic moves from explicit workflow diagrams to implicit agent reasoning.
Murag, Anthropic technical staff, explained the distinction: "Unlike RAG, this relies on simple tools that let Claude manage and read files from a filesystem. Skills can contain an unbounded amount of context to teach Claude how to complete a task or series of tasks."
Competing AI Operating Systems
The convergence of AgentKit and Skills within a week isn't coincidence, it's competitive positioning for the next phase of AI adoption. Both companies recognize that owning the execution layer matters as much as providing the best models.
The strategic question facing every player:
For LLM providers: Can you build comprehensive enough ecosystems including integrations, execution environments, and developer tools to become full-stack platforms rather than just model APIs?
For integration platforms: Can you evolve fast enough to orchestrate agentic workflows while maintaining the integration breadth that's your moat?
For enterprises: Do you standardize on one LLM provider's platform and accept lock-in, or maintain platform-agnostic orchestration at the cost of shallower integration?
The answers will determine whether we see competing AI "operating systems" or interoperable layers where models, orchestration, and integrations remain separable concerns.
The future likely includes multiple winners serving different needs, but all with stronger platform boundaries than current integration-agnostic tools. The trade-off: deeper capabilities within platforms versus easier portability across them.
For developers and enterprises, this means thoughtful platform selection becomes more critical. Choosing a primary LLM provider increasingly means choosing a complete execution environment, not just a model. The decision carries implications for tooling, workflows, and organizational practices, much like choosing a cloud provider or mobile platform.
For the AI industry, it validates that infrastructure and ecosystem matter as much as model capabilities. The competition goes beyond just "who has the best reasoning" to "who provides the most complete environment for getting work done with AI."
In motion,
Justin Wright
If Anthropic's Skills make Claude self-contained and OpenAI's AgentKit makes ChatGPT self-contained, each with their own integration layers and execution environments, are we heading toward a world of competing AI "operating systems" where portability between platforms matters more than raw capability within them?
I am excited to officially announce the launch of my podcast Mostly Humans: An AI and business podcast for everyone!
Episodes can be found below - please like, subscribe, and comment!