The most worth-watching trend this week: the AI Agent middle layer is being cleared out in batches.
Star projects that were still raising money last year have their code repos set to archived this year. Jasper dropped from a $1.5B valuation to a fire sale, and a16z-backed Yupp.ai shut down within half a year of launch. This isn't an "AI winter" urban legend-it's the inevitable "middle-layer crushing" that happens when platforms evolve.
Update 1: The 5 "Deaths" of Vertical AI Agents Are Now Summarized
A postmortem hot in the community this week categorized the AI Agent middle layer being cleared into 5 types:
- Shell-death: Just wrapped an API in a UI layer-once the platform upgrades model capability, product value goes to zero.
- Sync-layer-death: Core function is "move data from platform A to platform B"-meaningless once the official side launches native integration.
- Translation-layer-death: Relies on prompts to translate natural language into SQL/API calls-directly eliminated after models build in function calling.
- Prompt-tool-death: Sells "prompt optimizers," but models increasingly need no complex prompts.
- Workflow-assembly-death: Drags a few API nodes into a workflow-no one cares after the platform launches a native Agent builder.
Real impact on you: If you're using Coze/Dify to build simple workflows like "auto-write weekly reports," realize this is entry-level practice, not a commercial moat. What's truly valuable is the three-layer moat of private data + workflow embedding + evaluation power.
Update 2: What Do the Surviving Agents Look Like?
Matching the "clear-out" is another set of signals-some Agents are doing fine. Their common traits:
- Compliance moat: Medical, legal, finance-no matter how strong the model, it can't get the licenses and data.
- Private-data moat: Your data is unique, the model can't get it (e.g., an enterprise's internal case library).
- Deeply embedded in org workflows: Not a tool but part of the enterprise process-replacement cost is extreme.
- Holding evaluation power: Not a "tool that does things," but the one defining the standard of "what counts as done well."
Hands-on advice: Give your project a checkup with these 5 questions:
- If OpenAI ships this feature built-in next month, what do you have left?
- What non-transferable asset would users lose by leaving you?
- Is your data source something the LLM can also get on its own?
- Is your product a "necessary link" in the user's workflow?
- If the API rate-limits to 100 calls/day, would users go crazy?
If most answers wobble, what you're building isn't a moat-it's helping the platform dig an irrigation channel.
Update 3: The Engineer's "Barnacle" Survival Strategy
An inspiring metaphor this week: don't be a "lobster" (crawling fastest and most visible on the seabed-dies first when the environment changes), be a barnacle (parasitizing a giant whale, firmly anchored in a tiny niche, making it so the whale can't shake you off even if it wants to).
Concrete tech choices:
- Use LangChain or Dify to connect private knowledge bases and internal APIs, not just call public models.
- Make the Agent a scheduled task embedded in the user's workflow (cf. Ahrefs' SEO Agent: runs weekly, directly embedded in the user's SEO process).
- Keep data local, get industry compliance certifications-make the platform "want to eat it but can't."
Recommended resource: ai-agents-from-zero on GitHub has code from env setup to enterprise deployment-suitable for typing through to build intuition.
One-line This Week
What the platform wants isn't your feature, it's your position. Don't mistake your feature for your position.
Tools change, models upgrade, but "being an irreplaceable rivet at the user's throat" is a cross-cycle idea.