In brief: Focus is shifting from isolated AI models through agent harnesses to feedback loops as a core product, with signal design and evaluation mechanisms being critical to enable agents to act without constant human bottlenecks.
Startup Introspection, founded by former xAI employees, is building infrastructure for self-improving AI agents. At its core is the idea of feedback loops in which agents maintain and optimize the system itself, rather than merely executing external tasks.
Roland Gavrilescu, co-founder and CEO of Introspection, presented a new concept at the AI Engineer World’s Fair: autoresearch as an outer loop in which agents maintain and improve the primary system itself. They leverage feedback signals, automated evaluations, and human input as drivers for continuous improvement. Gavrilescu and co-founder Julian Bright previously worked at xAI, where they focused on agent infrastructure and cloud agents. They recognized a new agent application pattern that could not be fully pursued at xAI and founded Introspection to create specialized tools for it.
Gavrilescu outlines three core patterns for productive autoresearch systems: First, the feedback loop itself becomes the product—not the underlying AI architecture. For agents to take on more tasks without producing lower-quality output (so-called “slop”), the right feedback mechanisms must be defined. Second, Introspection introduces the concept of agent recipes—a format that functions like a translation of data recipes from AI training. An agent recipe documents how a harness works with different models, which evaluations are performed, what automated judge systems exist, and what human knowledge has been codified. This format is intended to be vendor-neutral and portable, similar to a research lab that the operator controls themselves.
As a third pattern, Gavrilescu identifies the question of the optimization target: How can a system become both more capable and more cost-efficient simultaneously? Companies like Cursor and Cognition have already proven such products work. The next step is to make them more accessible, faster, and cheaper—gradually distilling capabilities of frontier models into systems that an enterprise controls itself and can optimize for its specific environment. The central challenge remains: agents must first learn from humans before they can make autonomous improvement decisions.
Source: www.latent.space · Published July 2, 2026
Lumi AI News — AI-assisted curation pursuant to Article 50 EU AI Act. Paraphrase and classification by Lumi News Pipeline v1.7.2.