As AI software development transitions from developer laptops to cloud infrastructure, a fundamental shift in engineering roles is taking place. While product engineers focus on shipping customer-facing features, factory engineers focus on performance-tuning and optimizing the underlying agentic systems that produce software.
From Local Agents to Software Factories
Running local agents (like Claude Code or Codex) on individual laptops creates severe operational bottlenecks: human gating, unmanaged costs, and security risks from ungoverned local MCPs. A cloud software factory replaces this setup with an automated, sandboxed environment defined entirely in code.
The Core Responsibilities of Factory Engineering
Factory engineering treats the software development pipeline as open, measurable infrastructure. Key responsibilities include:
Factory as Code: Declaring agents, MCPs, repositories, and workflows in version-controlled code.
Cost & Workflow Metrics: Tracking operational telemetry—such as cost per PR, shipping velocity, and human interventions—to optimize spend (e.g., reducing cost per PR from $55 to $15).
LLM-as-a-Judge Scoring: Running automated online evaluations across sampled agent runs to identify failure modes like redundant unit test writing.
Closed Self-Improvement Loops: Deploying specialized agents that analyze recurring evaluation failures and submit pull requests to update the factory’s own codebase.
Custom Benchmarking & Model Routing: Replaying real team task histories across model matrix tests to plot cost-versus-correctness Pareto curves, guiding intelligent model routing.
By transforming development from individual prompts into a self-improving engine, factory engineering enables teams to safely scale software delivery.
source: https://www.youtube.com/watch?v=x-PK3yx_MJg