President & Head of AI at Replit, pioneering autonomous agents that create and deploy software. Previously Head of Applied Research at Google Labs and Google X, contributing to PaLM’s coding capabilities. Stanford instructor who began applying Transformers to code in 2018.
Architect of Autonomous Software Creation
Michele Catasta has emerged as a leading architect of the paradigm shift from AI copilots to fully autonomous agents. His work at Replit demonstrates that AI must move beyond support roles to true autonomous task completion, where agents handle all technical decisions within well-defined scopes.
Current Work
As President & Head of AI at Replit, Michele architected and led the launch of Replit Agent v1 (September 2024) and v2 (February 2025), driving over two orders of magnitude in company revenue growth. His innovations focus on:
- Autonomous Agent Architecture - Context management through persistent filesystem storage, sub-agents to prevent “context pollution,” and core loop orchestration deciding parallelism on-the-fly
- Autonomous Testing - Breaking the feedback bottleneck that accumulates small errors and overcomes frontier model “laziness”
- Scoped Autonomy - Agents make all technical decisions within specific, well-defined tasks while users maintain control over what matters to them
- Irreducible Runtime - Maximizing natural task complexity progression: plan → implement → test → loop
Michele co-authored the Replit AI Manifesto and led development of Building LLMs for Code Repair. He teaches Vibe Coding 101 with Andrew Ng on DeepLearning.AI.
Background
Before Replit, Michele served as Head of Applied Research at Google Labs and Google X, focusing on Large Language Models and contributing to the coding capabilities of PaLM and PaLM 2. Earlier, he was a Research Scientist and Instructor at Stanford University’s AI Lab (SAIL), where he pioneered applying Transformer architectures to source code starting in 2018.
He earned a Ph.D. in Computer Science from EPFL and co-founded Sindice.com, the largest Semantic Web Search Engine, which evolved into Siren (an investigative intelligence platform securing $15M+ funding). His research has earned 16,000+ citations on Google Scholar.
Beyond his operating roles, Michele is an active angel investor with 100+ investments in the AI space and serves as advisor to Bessemer, Coatue, Fellows Fund, and Obvious.
Philosophy on Agent Autonomy
Michele’s approach challenges the AI copilot paradigm:
Scoped autonomy over supervised assistance - Agents should make all technical decisions within their defined scope, not just suggest options. Users maintain control only over aspects they care about.
Context architecture matters - Using persistent filesystem storage, code base understanding, and memory dumps to manage context. Sub-agents invoked with fresh context protect the main agent’s working memory from “context pollution.”
Parallel agents for user experience - Long run times aren’t satisfying for productive people. Core loop orchestration decides parallelism on-the-fly while handling complexity that non-technical users shouldn’t face (like merge conflicts).
Testing breaks the bottleneck - Autonomous testing prevents accumulation of small errors and overcomes frontier model laziness, enabling true autonomous operation for extended periods (200+ minutes).
Replit
Replit is an AI-powered cloud development platform with 22+ million creators and 500,000+ business users. Replit Agent enables anyone to build professional web applications using natural language, with fully autonomous capabilities that test, fix, and deploy applications independently.
Conference Appearance
Event: AI Engineering Code Summit 2025 Date: November 20, 2025 Time: 9:25 AM Session: Autonomy Is All You Need
Michele presented the architectural principles behind autonomous agents, focusing on context management, sub-agent design, parallel orchestration, and autonomous testing. He introduced concepts like “painted doors” (wireframes) and “accumulations of whatevers” (frontier model laziness), demonstrating how proper agent architecture enables extended autonomous operation while maintaining user control over what matters.