The Post-AGI Era

This argues that AGI has crossed a practical threshold: models like Astra now reason across domains and perform complex tasks with real autonomy, which lets organizations rethink work design rather than merely using AI as a more powerful tool. The crucial distinction is between "good enough" AI that augments existing human workflows and AGI that enables those workflows to be decomposed into systems of collaborating autonomous agents. A new company form becomes possible: organizations built from the ground up around recursive self-improvement (RSI). An example described is Infinite Ascent, which implemented a harness and trading agents, added reviewer agents and a meta-harness to reinforce high-performing behaviors while preserving creativity, created an internal IT auditor to improve infrastructure, and plans to post-train models on its own performance data so the entire stack can evolve.

A persistent asymmetry exists between frontier model capabilities and the pre-AGI architectures most organizations still use, creating an opening for first movers who can design post-AGI companies and accumulate data and iterative advantage. The frontier model layer is mostly closed, producing concentration risk, information asymmetry, and provenance uncertainty (illustrated by the Navier-Stokes debate), and it prevents full RSI because organizations cannot push learnings into the core model. True self-improvement therefore requires control of the model layer - post-training, specialization, and continuous reshaping - so agents that start from the same base can diverge and deepen competence in different domains.