Descriptions:
AI Jason examines what Y Combinator is calling “self-improving companies” — AI-native organizations where agents handle internal operations autonomously, capture outcomes, and iterate on their own tooling without human direction. YC reports that companies in the current batch are achieving five times more revenue per employee compared to 18 months ago, with at least one company’s agents autonomously writing 45 of its own tools.
The video contrasts open-loop AI workflows (task completion without feedback) against closed-loop architectures where agent outputs continuously feed back into an intelligence layer. Drawing on a framework from YC’s Diana, the design breaks into five components: data ingestion, a policy layer encoding SOPs, tool access integrations, quality gates (human or AI evaluators), and a learning mechanism that improves future runs. The recommended starting point is a memory layer split between a temporal log of past actions and a procedural knowledge base that evolves into agent skills, paired with cron jobs for recurring execution.
Two case studies ground the theory. An SEO automation loop using Google Search Console, Ahrefs, and Google Analytics tripled organic traffic within one to two months for one practitioner. An autonomous ad campaign agent tested 10 different creative formats over several weeks, discovered that low-production “ugly” assets outperformed polished ones, and generated 243 leads on a $1,500 budget. The video also introduces an open-source agent skill framework for long-horizon tasks, available via the description link.
📺 Source: AI Jason · Published June 02, 2026
🏷️ Format: Workflow Case Study
