The loop is the product, not the model

Introspection co-founder Roland Gavrilescu told AI Engineer the next frontier is not a better model or a better harness but the loop that keeps improving itself.

He and co-founder Julian Bright left xAI a few months ago after working on agent infrastructure and cloud agents there to build a standalone company around always-on, long-horizon tasks. The talk frames a 2026 blueprint: trust the loop, distill what it learns, and measure valued work per watt.

The punchline came early.

Gavrilescu traced the arc from RL-trained reasoning to harness engineering to loops that run without touching code, then pointed to the first viral example that was not a coding agent at all. An engineer named AJ used what was then called Clawbot, now OpenClaw, to scrape inventories, pull Reddit prices, ping dealers and play PDF quotes against each other until the price was verifiable and the car could be locked in. That pattern, what he called the loop is the product, was already selling out Mac minis when developers started buying dedicated boxes to keep agents always on.

That history has a longer tail. Gavrilescu rooted the idea in OODA loops, the observe-orient-decide-act cycle coined for US Air Force fighter pilots in the 1970s, and argued models have been trained with that cadence in mind: call tools, take observations, act again. Put strong signals and verifiable work at the end and you get workers like Claude Code agents. The quality of the signal and the verifier decides whether the loop actually succeeds. Feed the artifacts of one loop back as the signal for the next and you get continuous improvement.

The second move is how to keep what the loops learn. Gavrilescu argued system distillation is the moat, and that the missing piece is a portable recipe for AI systems. Where RL teams learned to iterate data recipes to curb hallucinations and reward hacking, most agent teams have no equivalent for harnesses, evals, models, tools and environment. An agent recipe, in his framing, is a versioned, provider-agnostic package that lives in a Git repo and encodes not just prompts and skills but judges, profiles, resources and the human judgment that tuned them over time. Introspection builds that approach on the Pi harness and Harbor for evals, with an early release called pi.recipes that extends what skills were in 2025 into a full recipe that can be owned by the customer and managed by agents. Pi, in a separate interview, he likened to Linux: not meant to run vanilla, but to be extended through separate extensions and configuration so the same agent can be reconstituted elsewhere.

The practical flow he walked through was a talent-sourcing agent. Traces reveal a pattern, like chasing big-tech employees instead of hidden gems on GitHub. Patterns cluster those frustrations. Calibration judges and evals codify the taste: did the trajectory hit the wrong population, do humans agree hidden gems is the right call, and can an agent generate that judge with a human calibrating rather than hand-coding every test. Once calibrated, the fix becomes a recipe candidate, a diff that can be checked offline and then proven with users via A/B tests or multi-armed bandits before promotion. Repeated behavior becomes skills and prompts, failure patterns become evals, user frustration becomes harness memory. Do that over and over and the recipe becomes the taste of the maker, portable to others.

The third metric is valued work per watt. Gavrilescu pointed to Cursor and Cognition as the template: ship the best product, then the best evals for that product, then the best models distilled from both. Code got there first. Customer support, legal research and other verticals will follow the same curve, he said, but only by running systems in production to discover what frontier actually means before trying to make it economically viable. The open question is not whether the loop can be built but whether it can be made efficient enough that the price gap would pull users off a general harness like Claude Code. Introspection is pitching that as managed infrastructure for vertical SaaS: Git as the audit log, agents as mini researchers, humans as a tool that answers when asked and teaches taste until the system can act more autonomously. It is still early, and the whole argument rests on getting calibration right without letting outer-loop token bills run unchecked.