Introspection co-founder outlines ‘autoresearch’ patterns for self-improving agent systems

Roland Gavrilescu, co-founder and CEO of Introspection, describes autoresearch as a feedback-driven outer loop where agents help maintain and improve the primary system using signals, evaluations, and human input. The goal is to design mechanisms that let agents make architectural decisions and iterate without constant human bottlenecks.

Gavrilescu outlines three production patterns for autoresearch at the AI Engineer World’s Fair. First, the loop itself becomes the product, shifting focus from models to harnesses to closed-loop systems that can improve without generating more low-quality output. Second, he proposes the concept of an agent recipe—a portable container that bundles evals, judges, signal processing, and human expertise into a format agents can iterate on. Recipes aim to encode how components evolve, including failures that lead to new evals or model choices. Third, he emphasizes optimizing systems to become both better and cheaper over time, drawing parallels to how companies like Cursor and Cognition have operationalized agentic coding assistants.

Introspection’s Pi framework is compared to Linux, serving as a portable, extensible agent harness that separates the loop from its extensions and configurations. The company positions itself as a downstream distributor—akin to Red Hat—providing managed infrastructure to make agent loops reliable, cost-controlled, and secure in production environments.

Gavrilescu stresses that humans remain integral to the loop, with agents trained to query humans early in their operation and gradually reduce reliance as they accumulate preferences. This mirrors how new employees learn an organization’s norms before acting autonomously. The company is focusing on vertical agents beyond coding, targeting industries seeking secure, provider-agnostic deployment of agentic systems.