Tools

Introspection outlines practical patterns for self‑improving ‘autoresearch’ agent loops

Roland Gavrilescu, co‑founder and CEO of Introspection, described at the AI Engineer World’s Fair how autoresearch — an outer loop of agents that studies and improves a primary system — can be made reliable and deployable.

Introspection outlines practical patterns for self‑improving ‘autoresearch’ agent loops

Roland Gavrilescu, co‑founder and CEO of Introspection, discussed autoresearch today at the AI Engineer World’s Fair. He defined autoresearch as an “outer loop” of agents that study, maintain and iteratively improve a primary system using evaluation signals, feedback and human input.

Company background

Gavrilescu said Introspection grew out of work he and his co‑founder Julian Bright did on agent infrastructure and cloud agents while at xAI. They identified a new agent form factor worth exploring in depth and decided to build a company focused on the operational and product needs of self‑improving systems.

Three core patterns for autoresearch

Gavrilescu presented three design patterns he considers foundational for bringing autoresearch into production:

  • The loop is the product: attention is shifting from models to harnesses to loops. The central design problem is defining feedback mechanisms that let agents shoulder more work without introducing excessive noise or waste.

  • Agent recipes: a larger container that assembles the components necessary to encode human expertise — evals, judges, signal processing and the information fed back into the loop. Recipes aim to be a portable, provider‑agnostic format that records how a system evolved.

  • Optimize for better and cheaper: systems should become both more capable and less expensive over time. Early products from companies like Cursor and Cognition show feasibility; the next step is wider accessibility, speed and cost reduction, and distilling frontier model capabilities into owned, customized systems.

What is an agent recipe?

Gavrilescu described an agent recipe as akin to an ingredient list plus an evolution log. Drawing parallels to data recipes used in model post‑training, an agent recipe records how a harness interacts with models, which evals and judges are used, what human expertise has been captured, and which failures spawned new evals. The recipe captures the process so that future engineers can understand not just the code but the decisions, mistakes and signals that led to the current state.

Inner loop vs. outer loop

Autoresearch separates an inner loop — the primary system interacting with users and doing work — from an outer loop: a separate system that studies and maintains the primary one. The engineering challenge is designing the outer loop so it makes progress on the right problems while controlling token and compute costs used to decide actions.

Pi as the Linux of agent harnesses and open building blocks

Gavrilescu compared the Pi framework to Linux: Pi is intended as a base for agent harnesses that can be extended rather than a one‑size‑fits‑all product. By separating the agent loop from extensions and configuration, Pi makes agents portable. Introspection seeks to combine that extensibility with recipes and open‑source components so customers can run evolving, portable agents.

Making loops reliable in production

Introspection’s product targets the point at which teams are ready to move to production. That requires knowing the necessary infrastructure to run loops, control costs and maintain security. Introspection’s managed infrastructure is intended to provide the kinds of capabilities available in frontier AI labs in a deployable product for other companies.

Humans remain part of the system

Gavrilescu emphasized that humans are an intentional part of early autoresearch loops. Humans provide signals and act as tools: agents can be trained to use an “ask a human” action. In initial iterations agents may frequently query humans and learn their preferences; over time those preferences are captured and the agents can become more autonomous. Gavrilescu likened the process to a new hire who asks many questions initially and later makes independent decisions.

Focus on vertical markets and Git‑based workflows

Introspection is concentrating on vertical agents. While coding agents have demonstrated clear value, deploying agents in non‑coding, industry‑specific domains is now the focus. Customers want secure deployments, data ownership and freedom from lock‑in with providers like OpenAI or Anthropic. Introspection’s approach uses open‑source building blocks and a Git‑based environment where repositories and commit histories act as the audit log for agent work.

Practical advice for engineers experimenting with autoresearch

Gavrilescu offered three concrete recommendations:

  • Invest in signals: decide which feedback matters and build mechanisms to filter and surface the signals agents should act on.
  • Control cost: prevent runaway compute or token bills by managing loop efficiency and spending.
  • Follow research: study harness patterns and data recipe ideas from research labs and adapt those practices into your product.

His broader vision is to help product organizations become small research labs where agents operate as miniature researchers, extracting tacit human knowledge and gradually increasing automation rather than attempting to automate everything from day one.