ChatGPT Apps as Distribution & Open Agent Skills Ecosystem

Lenny Rachitsky - ChatGPT Apps: The Next Great Distribution Channel#

Source: https://www.lennysnewsletter.com/p/chatgpt-apps-are-about-to-be-the Credibility: High (detailed analysis with named enterprise partners and concrete examples)

What happened: Lenny published a deep analysis arguing ChatGPT apps represent a rare, decade-scale distribution opportunity—comparable to the App Store (2008) or early SEO. Major companies have already launched: Adobe, DoorDash, Canva, Figma, Booking.com, Coursera, Expedia, Spotify, and Zillow.

Key insights:

Discovery is contextual, not search-based: ChatGPT "automatically suggests apps based on a user's conversation." Users don't need to know your app exists—the model matches intent to solutions. This is fundamentally different from app stores or search engines.

Architecture (3 components):

1. Conversation interpretation (understanding user intent)
2. Backend tools (functions ChatGPT can invoke)
3. React-based UI widgets (rendered in a secure sandbox)

Tool optimization as "AEO" (Agent Experience Optimization): Tool names and descriptions function like a new SEO layer. The better your descriptions match user intents, the more likely ChatGPT surfaces your app.

Low barrier, massive reach: Simple apps require only "a few weeks to build." Solopreneurs can compete without marketing budgets because distribution is embedded in the conversation.

Why it matters for PMs: This changes distribution economics. Traditional channels require marketing spend; ChatGPT apps get surfaced when the model thinks they're relevant. For PMs planning product strategy, the question isn't just "should we build this?"—it's "where does this get discovered?" If your product solves problems people describe in natural language, a ChatGPT app may be your highest-leverage distribution investment.

Critical questions:

• How does ChatGPT decide which app to surface when multiple are relevant?
• What's the competitive moat once everyone has ChatGPT apps? Back to product quality?
• OpenAI controls the discovery algorithm—how much platform risk is acceptable?

Action you could take today: Evaluate your product's core use cases. If users frequently describe their problem in natural language (vs. clicking through menus), prototype a ChatGPT app. Focus on tool descriptions—they're your new meta tags.

Vercel - Open Agent Skills Ecosystem: skills.sh#

Source: https://vercel.com/changelog/introducing-skills-the-open-agent-skills-ecosystem Credibility: High (first-party announcement, CLI available now)

What happened: Vercel launched an open skills ecosystem for agents—standardized packages that extend agent capabilities across platforms. Skills work with Claude, Cursor, GitHub Copilot, Gemini, and others.

How it works:

npx skills add <package>          # Install a skill
npx skills add vercel-labs/agent-skills  # Example

Key capabilities:

• Discovery: Browse skills by category and popularity at skills.sh
• Cross-platform: Skills work across multiple agent systems
• Analytics: Track usage stats and installs across the ecosystem

Why it matters for PMs: This is Vercel positioning skills.sh as "npm for agents." If successful, it creates a network effect: developers build skills to reach users across multiple agent platforms; users get more capabilities without waiting for each platform to build native features. For PMs building agents, the question becomes: which capabilities do you build vs. install from the ecosystem?

Critical questions:

• How do you ensure skill quality and security when installing third-party packages?
• Cross-platform compatibility sounds great, but how different are the actual integrations?
• Who controls the ecosystem governance as it scales?

Action you could take today: Visit skills.sh and browse existing skills. If you're building an agent, identify which capabilities could be installed vs. built. If you're building tools, consider whether publishing as a skill could accelerate distribution.

LangChain - Agent Behavior Analysis: From Traces to Insights#

Source: https://www.blog.langchain.com/from-traces-to-insights-understanding-agent-behavior-at-scale/ Credibility: High (production experience with concrete problem statement)

What happened: LangChain published on the challenge of analyzing agent behavior at scale. The problem: teams collect hundreds of thousands of traces daily, but "it's impossible to read and summarize 100,000 traces at any human scale."

Why agents are different from traditional software:

1. Non-determinism: Same input produces different outputs each run
2. Prompt sensitivity: Tiny input changes create dramatically different results
3. Unbounded input: Natural language creates infinite intent possibilities

Why traditional analytics fail: Product analytics (clicks, sessions, funnels) assume bounded, structured interactions. Agent conversations are unstructured. Metrics show what changed but not why. Online evaluators only answer questions you've already thought to ask.

The solution: Clustering-based insights: LangSmith's Insights Agent uses clustering to automatically discover behavioral patterns across thousands of traces without predefined criteria. It surfaces unexpected patterns, failure modes, and usage themes organically.

Why it matters for PMs: If you're running agents in production, your debugging tools are probably wrong. Traditional monitoring shows metrics; agent debugging requires understanding conversation patterns. The clustering approach—discovering themes you didn't anticipate—is critical because agent failures emerge in production, not testing.

Critical questions:

• At what scale does manual trace review become impossible? (Apparently <100K/day)
• How do you prioritize which discovered patterns to investigate?
• What's the latency from pattern discovery to production fix?

Action you could take today: If you're running agents in production, estimate your daily trace volume. If you're manually reviewing traces, calculate how much you're actually seeing as a percentage of total. If <5%, you need automated pattern discovery.