Michele Catasta

Michele Catasta is the president and head of AI at Replit, where he leads development of Replit Agent, software designed to let people without programming expertise build working applications. His career spans semantic search, graph-machine-learning research, Google’s PaLM models, and increasingly autonomous coding systems.

Catasta co-founded the semantic-web search engine Sindice, earned a computer science doctorate at EPFL, and conducted research at EPFL and Stanford, including a machine-learning postdoctoral fellowship under Jure Leskovec. He co-authored the Open Graph Benchmark, which established standardized datasets and evaluation protocols for graph machine learning.

At Google X and Google Labs, he led applied research on language models and source code and contributed to PaLM and PaLM 2. He joined Replit as vice president of AI in June 2023, laying out an agent-centered vision for software development in which models use programming tools, observe execution, and recover from errors.

His early Replit work included Replit Code v1.5, a compact, openly released coding model trained on filtered, permissively licensed data and optimized for practical inference. In 2024, he served as principal investigator on research into automated code repair and led the initial launch of Replit Agent. Later generations expanded the system’s ability to build, test, and coordinate substantial software projects.

• Autonomy without technical supervision. Catasta defines meaningful autonomy by the amount of useful work an agent completes without requiring users to make engineering decisions. Users should control what they want built; agents should handle implementation. Long runtimes alone prove nothing.

• Verification against convincing failures. Generated applications often look finished while buttons, databases, or integrations remain broken. Catasta combines static analysis, execution, API checks, logs, and Playwright-based autonomous testing to detect these failures. Agent-generated browser tests also become reusable regression coverage.

• Focused context and agent-directed orchestration. Plans and intermediate state can live in files instead of filling a model’s context window. Specialized subagents isolate noisy work such as browser testing, return concise results, and make parallel execution practical without requiring nontechnical users to coordinate tasks or resolve conflicts.

• Continual learning for agents. Catasta argues that application teams can improve agents even without owning the underlying model weights. Better harnesses, evaluation, production feedback, verification, and continual learning turn real-world failures into opportunities to improve software-building systems.