Driving · 2026-09-30
Towards an AI Software Factory for Data Systems
Anna Pavlenko, Bogdan Crivat, Brandon Haynes, Carlo Curino, Fotis Psallidas, Jaro Slawinski, Johannes Freischuetz, Laura Pereira Sanchez, Markus Weimer, Mathieu Demarne, Matthias Jasny, Mauktik Gandhi, Max Bovykin, Mirco Milletari, Purbasha Ghosh, Qiushi Bai, Raghu Ramakrishnan, Rahul Pandita, Sergiy Matusevich, Shivaram Venkataraman, Subru Krishnan, Md. Tareq Mahmood, Tiemo Bang, Venkatesh Emani, Xuan Zhao, Yiwen Zhu
Auto-summarized: this summary was generated by a language model from the paper’s text and has not been reviewed by an editor. Check the paper before relying on it. Benchmark numbers appear only after a human has verified them.
TL;DR
The authors propose an AI Software Factory for Data Systems that automates SDLC stages (Targeting, Coding, Reviewing, Ops) using a World Model and an evolutionary coding system (Darwin) wrapped in a self-improvement loop. They report scaled Microsoft deployments (tens of repositories) yielding 3× engineering efficiency vs agentic coding and up to 22× token efficiency, plus OSS and production evidence.
Why it matters
The paper targets the “Amdahl’s law effect” where large coding improvements do not translate into end-to-end SDLC throughput. It argues that for Data Systems/ECTs, adding decision tracing via a World Model and evolutionary agentic search can accelerate multiple SDLC stages and provide an SDLC-scale learning loop using decision traces to update the World Model and fine-tune model weights.
Method
- World Model: a metadata substrate that mediates telemetry/profiling, stores evidence, supports provenance and access control via purpose-specific viewpoints, and evolves as evidence accumulates.
- Hotspan targeting agent: transforms optimization goals into measurable objectives, enumerates targets from telemetry/code signals, ranks by ROI, and packages selected targets as Evolutionary Coding Tasks with evaluation harnesses (tests/benchmarks and counterfactual simulators).
- Darwin: an evolutionary agentic swarm for Evolutionary Coding Tasks that uses agentic memetic variation (code–test–debug–reflect cycles in sandboxed sessions) with hypothesis data substrates for guided exploration; integrated into an SDLC learning loop.
Limitation
Open Challenge: Generalizing World Model: supporting general purpose world modeling and update is a hard and open problem.
Abstract (from arXiv)
AI-assisted coding tools deliver significant acceleration of coding, but only limited impact across the end-to-end software development lifecycle (SDLC)--an Amdahl's law effect! In this paper, we discuss our progress towards building an AI SW Factory that accelerates all the stages of SDLC-Targeting, Coding, Reviewing, and Ops. The AI SW Factory produces a metadata exhaust that enables self-improvement by fine-tuning model weights and updating our World Model (a rich data substrate). We focus on Data Systems and the important class of Evolutionary Coding Tasks (i.e., those with a measurable objective to hill-climb) and report on 1) scaled deployments at Microsoft (tens of repositories) leading to 3x engineering efficiency above agentic coding and up to 22x token efficiency, and 2) several open challenges.
Related papers
- Lookahead-R: Budget-Aware Tool Retrieval via Execution-Centric Planning
- SkillWeaver: Agentic Exploration over Neural Interaction Skills for Scalable Robot Data Generation
- Explore, Execute, Evolve: A Skill Acquisition and Reuse Loop for Embodied Agents
- The GUI Is Not the State: Diagnosing State Aliasing in GUI World Models
- Transformers are Sample-Efficient World Models