Roland Gavrilescu is the co-founder and CEO of Introspection, building infrastructure for self-improving AI systems. His work centers on what happens after an agent launches: turning mistakes, expert judgment, and customer feedback into improvements that a company can inspect, preserve, and reuse.
From learning algorithms to deployed agents
Gavrilescu began exploring deep learning while studying electrical engineering in London. A Facebook-sponsored PyTorch scholarship gave him a practical introduction to the field, and he organized a local meetup where fellow participants worked through the material together.
His subsequent open-source work spanned reinforcement learning and computer vision. In 2020, he contributed to Stable-Baselines3, including support for multiple discrete and binary observations and MultiCategorical and Bernoulli probability distributions for PPO and A2C. The work helped the reinforcement-learning library represent environments and action choices beyond a single discrete choice or continuous control. His flaxvision project brought selected torchvision neural-network models into JAX and Flax; its initial development began during Google Summer of Code under Avital Oliver’s mentorship.
Gavrilescu later worked on AI agents at Superhuman and on agent infrastructure and cloud agents at xAI. At xAI, he met Julian Bright, with whom he founded Introspection. Their founding story connects the move to a practical ambition: making the machinery for improving deployed agents available to other businesses.
That ambition extends learning beyond model weights. A capable model can still perform poorly because of its tools, instructions, integrations, or execution environment. Gavrilescu’s November 2025 introduction to Introspection proposes an adaptation layer that examines production traces, diagnoses failures, creates evaluations, and proposes changes across the AI system.
Making experience reusable
Gavrilescu argues that an agent’s work should supply the evidence for its next improvement. The agent performs tasks in an inner loop; an outer loop studies the resulting traces and artifacts, identifies what should change, and tests those changes. Strong feedback signals guide the work, while verifiers check whether apparent success reflects a useful result. His approach to autoresearch makes this improvement process part of the product.
Several ideas give that process its shape:
Portable agent recipes: Gavrilescu calls the capture of accumulated lessons “system distillation.” Failure patterns become judges and evaluations; recurring useful behavior becomes skills and prompts. Introspection’s Recipes format keeps agent configuration, optional skills, tools, subagents, judges, and evaluations in editable source directories built on Pi. In its early Pi recipes release, the team combined the Pi harness, Harbor evaluations, and Git repositories so owners and agents could track what changed and why. The intended advantage is knowledge that remains under the customer’s control as models and providers change.
Human-calibrated evaluations: Gavrilescu uses “taste” to mean the expert judgment that determines what good work looks like in a particular setting. His talent-sourcing example shows why this matters: an agent may favor prominent engineers or large technology companies, while a recruiter wants less obvious candidates with strong public work. Execution traces expose the mismatch. Agents can implement evaluations to detect it, but people must calibrate the judgment those evaluations encode. A/B tests with users then test whether a proposed change deserves promotion, rather than treating agreement with the maker’s preferences as sufficient proof of improvement.
Earned autonomy: Gavrilescu wants agents to acquire organizational knowledge through interaction. An agent can ask targeted questions, preserve the answers, and require less supervision as it learns. His account of autoresearch connects that learning to the outer loop: accumulated knowledge helps the system decide how to improve the agent doing the primary work.
Valued work per watt: Gavrilescu’s measure of progress joins useful outcomes to the resources required to produce them. Evaluations and user experiments establish whether customers value the work; further improvements must make delivering it economically viable. Capability alone does not establish a sustainable service.
Introspection’s product direction extends this approach beyond coding agents to businesses whose specialists hold accumulated knowledge about difficult workflows and exceptions. The company aims to turn that expertise and production experience into customer-owned agents, training data, and specialized models.
Roland Gavrilescu’s blueprint for improving agents connects execution, human judgment and production experiments: capture what a system learns in portable recipes, then optimize for work users value at a cost that makes sense.
Build the feedback path as carefully as the worker: signals guide execution, verifiers check success, and artifacts inform the next improvement loop.
Preserve operating knowledge in versioned agent recipes: failures become evaluations, repeated behavior becomes skills, and human judgment remains part of the system.
Let humans calibrate the desired judgment and agents implement the evaluations. Then test whether users value that judgment before promoting a recipe change.
Establish that the output is valuable, then make its economics viable. Efficient execution only becomes a useful product advantage when customers want the work.
▶ Watch ↗