⚡️ The 2026 Software Factory Infra Stack — Zach Lloyd, Warp.dev
Summary
Zach Lloyd discusses the concept of the 'Software Factory,' defining it as a measurable, agentic system in the cloud designed to automate and optimize the entire software development lifecycle (SDLC). The factory is defined declaratively in code, allowing for the orchestration of various agents, skills, and Micro-Component Protocols (MCPs) across a product surface. Key benefits include centralizing agent conversations, enabling detailed cost and throughput tracking, and establishing a closed-loop system for continuous self-improvement and benchmarking.
Key takeaways
-
Defining the Software Factory
7:10
A software factory is not merely a collection of automations; it is a specific, measurable, and orchestratable system defined in code (e.g., a YAML file). It operates on a product surface and defines a workflow (triage, specking, implementation) that mirrors traditional development processes.
-
Closed-Loop Observability and Measurement
14:00
The factory centralizes all agent conversations and artifacts, providing a full trace of agentic coding activity. This data allows teams to track traditional metrics (DORA) alongside specialized metrics like cost per PR, throughput, and agent efficiency.
-
Self-Improvement and Benchmarking
23:52
The system supports continuous optimization through 'scorers' (online evals) that retroactively grade historical agent runs for failure modes (e.g., redundant tests). Furthermore, it allows benchmarking by replaying tasks against a matrix of different models and harnesses (e.g., comparing GPT-5 vs. GLM-535 cost/correctness trade-offs).
-
The Stack View of Factory Engineering
30:30
The ideal factory infrastructure is a stack: 1) Declarative definition of the factory in code (like Terraform), 2) A data layer to gather all agent conversations and artifacts, 3) A compute layer (pluggable sandbox/namespace), 4) An orchestrator to move work through stages, and 5) Multiple access points (Slack, Linear, GitHub) to integrate into existing tools.
More digests on Software Factories
Technical details
-
Cost Optimization
Warp initially incurred a cost of approximately $80 per PR using high-powered models (like Opus) for triage. Through internal tuning and testing, the cost was reduced to roughly $15–20 per PR, while simultaneously increasing throughput by 40%.
-
Agentic Workflow Tracing
900s
The system provides a full record of agent actions, including computer use screenshots, allowing diagnosis of exactly what the orchestration agent was doing during a task, regardless of the starting point (Slack, Linear, etc.).
-
Model and Harness Flexibility
1610s
The architecture is designed to be highly pluggable, allowing users to integrate various inference providers (Anthropic, OpenAI, Google) and different harnesses (e.g., Cloud Code, Codex, Pi), rather than being locked into a single vendor.
-
System Scale and Adoption
2000s
The factory approach is recommended for companies with significant scale (5 to 10+ engineers) that require a platform system to accelerate development, rather than for very small teams.
Mentioned resources
Channel & topics
Watch on YouTube · Back to latest
This independent, AI-assisted summary is provided for commentary and informational purposes. It may contain errors or omit important context. Please watch the original video for the creator's complete presentation. Video, thumbnail, and related copyrights belong to their respective owners.