The initial wave of generative AI was defined by the prompt: a human asks, and a model responds. But the industry is rapidly pivoting toward "agents"—systems designed not just to speak, but to act. Trellis AI, a graduate of Y Combinator’s Winter 2024 cohort, is now signaling its intent to push this paradigm further by recruiting engineers to build what it calls self-improving agents.
The technical challenge lies in the gap between execution and optimization. While current AI agents can navigate browser tabs or write code, they often struggle with the recursive logic required to fix their own errors. A self-improving agent implies a feedback loop where the system evaluates its own output against a set of objectives, identifies friction points, and adjusts its internal logic without human intervention.
This move by Trellis AI reflects a broader shift in Silicon Valley from raw model scale to architectural sophistication. As the low-hanging fruit of large language model training begins to vanish, the competitive frontier has moved to the "inner loop" of autonomy. The goal is to create software that doesn’t just perform a task, but learns the specific nuances of a company’s workflow through sheer repetition and self-correction.
With reporting from Hacker News.
Source · Hacker News
