⚡️ The 2026 Software Factory Infra Stack — Zach Lloyd, Warp.dev

Summary

Zach Lloyd describes Warp Factories as infrastructure for running coding-agent workflows—not just a collection of automations. A factory is defined in code for a particular product area, with its repositories, skills, MCP integrations, and workflow stages specified together. Work can enter through tools such as Slack, Linear, Jira, GitHub, an API, or MCP, then proceed through the configured workflow.

A central feature is observability: the platform gathers agent conversations, actions, artifacts, and other task records in one place. Teams can use this information to measure and improve agent work, including:

• Delivery: Time to release, PR throughput, and how often humans intervene.
• Cost: Cost by PR, employee, automation, or workflow stage.
• Quality: Code quality, verbosity, and unnecessary tests. Teams can configure custom checks, often using an LLM as a judge.

Lloyd argues that companies should move from judging coding agents by intuition to evaluating them with data. Warp’s benchmarking feature lets teams replay real tasks across different agents, models, or testing configurations, then compare cost and accuracy. Those results can inform model routing and other workflow choices.

The interview also covers factory self-improvement. Agents can analyze performance data and propose changes to the factory’s code—for example, adjusting triage to use a subagent in certain cases. Automatically replaying past tasks to validate proposed changes is a planned capability, rather than something the system already fully does.

Lloyd outlines the broader infrastructure involved: a version-controlled factory definition; storage for agent activity; integrations and access controls; compute environments such as Linux sandboxes or Mac virtual machines; configurable inference providers and evaluation systems; an orchestrator; and multiple entry points into the workflow. Defining the factory declaratively makes it possible to version changes, branch, test, and roll back—and potentially let agents propose improvements.

As an internal result, Warp reduced its reported cost from about $80 per PR to $15–20 by changing model selection and agent structure. Lloyd says PR throughput rose 40% in the prior month, with no increase in cancellation rate.

Lloyd sees the approach as most useful once a team reaches some scale—roughly 5–10 engineers or more—with growing startups and larger engineering organizations as likely users. Very small teams may be better served by individual coding-agent subscriptions; companies with highly specialized infrastructure may prefer to build their own systems. He frames the emerging work as two related disciplines: product engineering, which builds features, and factory engineering, which designs systems that produce software.

Reviews, Guides, and Demos

• Product walkthrough: The interview demonstrates how a Slack task can enter a factory and how teams can inspect agent activity, scoring, and performance data.
• Benchmarking example: Real company tasks can be used to compare agent configurations on cost and accuracy.
• Format: No formal review or step-by-step tutorial is provided; the episode is primarily an interview and platform overview.

Main Speakers

• Zach Lloyd — Warp founder and CEO, discussing Warp Factories and factory engineering.
• Swyx — Host and interviewer.