Why 90% of AI agent startups will fail, and the architecture that survives.
What the last year of agent demos taught us about durable products, and the layer underneath the ones that hold up.
Most AI agent products today are not products. They are prompt-orchestration layers sitting on top of large models, and prompts, personas, and many agent frameworks will commoditize quickly as the underlying models improve. We have seen the pattern before: prompt marketplaces that appeared in 2023 and disappeared inside a year, generic copilots that struggled the moment the foundation-model layer caught up. AI alone is not the product. Execution infrastructure is.
Capital is in the agent layer. Survival is in the layer underneath.
Figures cited in the point of view; validate against current market data before quoting.
agentic AI market in 2025, up from ~$5.25B in 2024, with projections crossing $52B by 2030
of agent projects Gartner estimates will be cancelled or fail to ship by 2027
raised by vertical AI startups in 2025 alone, as the verticals begin to separate
Five layers. The model is not one of the moats.
The durable platforms in production, GitHub Copilot in the developer workflow, Notion AI in the knowledge workspace, Figma AI inside design files, share the same shape. The moat is the vertical environment where the work happens. The model can be swapped; the environment cannot.
Context layer
A real understanding of the environment the agent operates in: the repositories, documents, data, and historical decisions. Most failures live here.
Tool execution layer
Agents have to interact with the systems where work happens, GitHub, CRM, analytics, cloud, internal platforms. Without it, the agent has nowhere to act.
Workflow orchestration
AI embedded inside execution pipelines, from idea to specification, implementation, testing, and deployment. The agent is a step in the workflow, not the destination.
Memory and learning
Systems that improve over time by learning from organizational patterns and past outcomes.
Governance and human oversight
AI proposes actions; humans approve execution where it matters. Auditable, reversible, explainable.
"A simple test: if removing the underlying LLM breaks your product entirely, you probably do not have a durable product yet. The model should be a replaceable engine, not the foundation of the business."
Get the full point of view
The complete essay, with the execution-operating-system examples and the market context, for internal circulation.
Building an agent, or the system it lives inside?
A senior practitioner will map the execution layer that makes an agent durable, not just a demo.