We’ve seen 3 distinct waves of app-layer AI startups since ChatGPT was released, each with more competition and consolidation. We’ve just entered wave #3 and the competitive dynamics are shifting in interesting ways:

Wave 1️⃣ was wrappers and clever prompts w/ polished UIs that were quickly outpaced as the underlying models improved.

Wave 2️⃣ was “AI-native” startups building custom data pipelines and embedded workflows, but many got squeezed when the platforms they depended on launched competing apps.

Now we're in Wave 3️⃣, the agentic era. Powered by RL, this wave is about agents that learn - optimizing workflows over time instead of relying on static prompts. Model providers are shipping these agents straight to users (OpenAI’s Agent Mode, Anthropic’s Claude Code), and the bar for “AI-native” is rising fast.

Startups are building RL-driven agents that observe entire processes. A sales agent might learn that following up two days after a demo performs better than one, or that fintech deals close faster with a sales engineer looped in early. These agents learn tacit knowledge that doesn’t live in traditional CRMs.

Big Tech controls massive historical datasets (Google with search data, Meta with social data). What they lack is visibility into the full chain of decisions, actions, and handoffs that define how actual work gets done.

Better base models like OpenAI’s improvements only enhance these agents’ ability to learn, and the stream of interaction data generated by users and agents working together is becoming the key proprietary asset in AI.