Agent recipes — portable, Git tracked records of the prompts, evals, and human corrections an AI agent loop has used — are one of three production patterns Introspection plans to lay out at AI Engineer World's Fair this week.
Picture the moment an AI agent catches a regression before users do: the prompt template drifted, the eval suite flagged a judge-output mismatch, and a maintainer agent files the fix back into the same loop that produced it. That kind of self-correcting cycle is the working definition of "autoresearch." It is a feedback-signal system where AI agents help maintain and iterate the primary system, not a single model magically improving itself in isolation.
The most durable part of that loop, the new company Introspection argues, is not the model at the center of it. It is the recipe that encodes not just a baseline but an evolution path: which signal produced which judge, what human expertise got embedded, and which model swap occurred along the way. CEO Roland Gavrilescu, who co-founded the company with Julian Bright after the pair worked on agent infrastructure and cloud agents at xAI, plans to lay out three production patterns at AI Engineer World's Fair 2026 in San Francisco this week. The three patterns: the loop itself as the product (a shift from models to harnesses to loops), "agent recipes" as a portable container, and an explicit dual objective of getting better and cheaper over time.
The recipe idea is borrowed wholesale from model post-training data recipes, the curatorial logs that capture how a model was tuned, then transplanted onto the agent loop. Where a post-training recipe encodes "we trained on this mixture and removed this bad sample," an agent recipe encodes "we added this judge, we captured this human's correction, we swapped from model A to model B and the eval moved by X." That is a richer audit trail than a model card, and it is the structural bet underneath the company. It also explains why Introspection is so explicit about being provider-agnostic. The recipe is the asset, not the underlying model. The entire stack is Git-based, with Git serving as the audit log for every signal and human edit.
Underneath the recipes sits Pi, the framework Introspection positions as the "Linux of agent harnesses". The pitch is separation of concerns: Pi runs the agent loop, while extension and config files are loaded in to spin up different agents. That gives a small team a way to share, version, and fork the harness layer the same way Linux lets distributions ship different user experiences over the same kernel. Introspection's role is the distribution and services layer on top of Pi, not the kernel itself. That positioning reads as a deliberate hedge against being mistaken for yet another agent framework.
There is a useful historical anchor. Andrej Karpathy's open-source AutoResearch project, a small Python tool that runs a propose-train-evaluate cycle on a single GPU, is the upstream reference point many readers will already have in mind. Introspection uses the same word for something bigger: an outer loop that spans signals, evals, judges, human input, and cost control, rather than a single-node ML experiment. The risk for the reader is conflating the two. The risk for Introspection is that the broader framing sounds like a rebrand of agent-CI or eval pipelines that already exist. The AI Engineer World's Fair program, which lists Autoresearch alongside Software Factories and Harness Engineering as headline tracks, suggests the category is being formalized even if individual companies have not shipped it at scale yet.
The interview itself reads like a practitioner's checklist dressed up as a pitch. Gavrilescu's engineer playbook: invest in signals first, control cost so an inefficient loop does not generate a thousand-dollar surprise bill, follow the model-trained harness research coming out of frontier labs, and treat the product organization as a miniature research lab. The "software factory vs orchestra" framing is the most provocative part. Build toward factory autonomy, but keep the human in the loop as a core component for extracting tacit knowledge that agents cannot generate on their own. The "ask a human" tool, the interview suggests, is used heavily in early loops and then fades as captured expertise gets baked into recipes.
What to watch next is concrete. The three production patterns from the AIEWF talk need a named mechanism each, including input, signal, and action, before this category is testable rather than slogan-shaped. Independent practitioner voices, especially on the eval and RLHF side, are the missing piece. Every public claim about Introspection's positioning today is CEO-attributed, and Cursor and Cognition are cited only as design inspirations, not customers. If AIEWF slides and a recording land this week, the engineering claims become checkable. Until then, "agent recipe" is a useful vocabulary for any team running production agent loops, a way to ask whether your own harness is encoding real institutional knowledge, or just burning tokens at a more sophisticated layer.