Autoresearch: The feedback loop behind self-improving agents

Roland Gavrilescu spent last year building agent infrastructure at xAI before deciding the interesting problem wasn't inside the lab — it was in how companies outside frontier labs could get similar self-improving systems. So he and co-founder Julian Bright left to start Introspection, a company betting that the next wave of AI products won't be defined by better models or slicker harnesses, but by feedback loops that quietly get better on their own.

The core idea, which Gavrilescu laid out ahead of his AI Engineer World's Fair talk this week, splits any agent system into two loops. The inner loop is the thing users actually touch — the coding assistant, the customer-support bot, whatever does the work. The outer loop is a second system that watches the inner one, gathers signals from evals, judges and human feedback, and nudges it toward better architecture and behavior. Autoresearch, in his framing, is really about designing that outer loop so it fixes real problems instead of burning tokens chasing noise.

To make that process portable, Introspection is pushing a concept it calls an agent recipe — borrowed loosely from the 'data recipes' AI labs use in post-training, which specify how much data from which domains goes into a model. An agent recipe does something similar for a live system: it tracks which evals were built, which judges were created, which human decisions got baked in, and which failures led to new checks. Gavrilescu's pitch is that raw code, even something as capable as Devin's codebase, is nearly useless without the history of mistakes and fixes that produced it. The recipe is meant to capture that history so another team, or another agent, can actually learn from it.

Underneath all this sits Pi, an open-source agent framework Gavrilescu compares to Linux — a base layer nobody runs unmodified, built to be extended with different models, tools and configurations. Introspection positions itself as something like a Red Hat for that ecosystem: managed infrastructure for cost control, security and reliability once a team is ready to put these loops into production, using open building blocks instead of locking customers into OpenAI or Anthropic.

Humans don't disappear in this picture, at least not at first. Early on, an agent is expected to lean on an 'ask a human' tool constantly, the way a new hire asks questions before they understand how a company actually operates. Gavrilescu argues you can't dump an organization's tacit knowledge into a markdown file and expect autonomy on day one — the system has to earn it by extracting real decisions from real people over time. That's also his answer to the 'orchestras versus factories' debate going around the conference: build toward the fully autonomous factory, but don't pretend you can skip the human-conductor phase first.