Emulated Founders Detail Data Engine for Autonomous AI Engineers

AI coding agents can generate pull requests, but running an enterprise software platform requires far more than isolated code edits. At the AI Engineer World's Fair, Joseph Wang and Sid Patlollu, co-founders of data lab Emulated, argued that training truly autonomous software engineers requires shifting from single-container code sandboxes to multi-node cloud environments that mimic entire tech companies.

The Founders Behind Emulated

Joseph Wang, CEO of Emulated, along with CTO Sid Patlollu, built their technical foundation in network infrastructure, distributed databases, and sandbox architecture. Their experience with mission-critical platforms highlighted a distinct capability gap in current AI models. While frontier large language models handle high-level application logic well, they routinely fail when confronting low-level infrastructure problems such as multi-version concurrency control, clock skew, or data corruption in distributed database engines.

Why Code Diff Benchmarks Fall Short

Popular benchmarks like SWE-bench Pro or Terminal Bench evaluate agents across 50 to 100 turns, resulting in pull requests containing a few thousand lines of code. Wang pointed out that this setup omits crucial engineering work. Real software engineering requires discovering context across stale tickets, interpreting customer feedback, executing gradual rolling deployments, and observing service behavior under live traffic load.

"The model capability gap is a data gap," Wang said during the talk. He emphasized that AI models do not regress when trained on higher-quality datasets, meaning current limitations stem directly from low-fidelity training setups.

Moving from Single-Node Sandboxes to Real Cloud Environments

To train models for real infrastructure duties, Emulated places entire software engineering environments into containerized setups. Agents face issues like lagging node learners, flapping nodes, and network partitions. Patlollu demonstrated a multi-node etcd consensus cluster simulation where agents must adjust cluster state while maintaining service uptime to minimize operational blast radius.

However, Wang explained that single-node deterministic simulations hit a wall when attempting to model cloud-scale platforms like Alphabet Inc. (NASDAQ:GOOGL) or AWS. Simulating real cloud platforms requires host provisioning, VPC setup, front-end APIs, authentication, telemetry, and billing engines.

The Enterprise Reality and Market Comparison

Building high-fidelity simulation environments for reinforcement learning remains one of the hardest problems in software automation. StartupHub.ai data rates platform capabilities in this domain with a 71/100 score, placing Emulated alongside major industry players such as OpenAI, which holds an 84/100 score, and Perplexity AI at 71/100. By focusing strictly on infrastructure workflows, Emulated aims to train agents that manage complete software lifecycles safely and reliably.