Zach Lloyd

Latent Space: The AI Engineer Podcast · ⚡️ The 2026 Software Factory Standard — Zach Lloyd, Warp.dev

• How Does Warp's Software Factory Improve Itself?

  Zach Lloyd says Warp's software factory can improve itself because the factory is defined as code. A second loop of agents watches the scores and looks for failures. On the triage step, those agents found 54 failure cases where the triager should have spawned a sub-agent to investigate. Because the definition is code, the agents can open a pull request against it and try to improve the factory, not only the product the team is shipping.

  Warp calls that loop self-improvement. The suggested change is not checked by replaying the old runs yet. The next step is to rerun those tasks and confirm the change would have helped. That loop is how a team starts engineering the system that writes the so

  ▶ at 14:36 Share Transcript

• Why Does Warp Stay Out of Token Reselling?

  Zach Lloyd says paying by the API is 5 to 50 times more expensive than buying a pile of subscriptions. Small companies he talks with want the factory but stay on a rotating collection of $200 subscriptions, and that rotation is the cheapest way. They are probably not running anything like 17 different subscriptions. He does not know whether the terms allow the rotation. The other benefits disappear, it does not scale, and past a certain company size it will not be allowed.

  Shawn Wang adds that a $200 plan includes $8,000 of API usage, and Lloyd agrees that is not money to ignore. Lloyd does not want Warp in the pure coding-agent business, because a competitive coding agent cannot be built o

  ▶ at 31:05 Share Transcript

• When Does a Software Factory Make Sense?

  Zach Lloyd says a software factory starts to make sense around five to 10 engineers. For a team of only a few people, the setup is overkill: use a Claude Code subscription or a Codex subscription and go faster. A larger team eventually needs a platform built to accelerate the work.

  The need runs up to very large companies, and the bigger the company, the more implementation work it takes. Standing the system up inside a company's own infrastructure is not purely plug and play, because the internal developer setup grows more complex with size. He does not know how much of that work Cursor takes on. For a fast-growing startup of 10 to 100 engineers, the fit is often self-serve. On what he cal

  ▶ at 27:15 Share Transcript

• What Did Warp First Pay Per Pull Request?

  Zach Lloyd says the factory approach at Warp first cost about 80 dollars per pull request. The team was using Opus on work that did not need it. The triage agent ran a high-powered model, and the agent setup spawned too many sub-agents instead of fewer agents, so a lot of the work was redundant.

  They ran internal tuning and testing with the built-in features and brought the cost per pull request down to about 20 dollars or 15 dollars. Over the last month, pull request throughput went up 40 percent. They also track the revert rate, and it has not gone up.

  He puts his confidence in observe, measure, tune. Watch how the agents spend money, change which models run and how many agents fire, the

  ▶ at 24:02 Share Transcript

• What Is Factory Engineering at Warp?

  Zach Lloyd says factory engineering means replaying a company's real tasks across models and harnesses so the team can see cost against correctness, and the target is the upper left corner: lowest cost, highest correctness. Warp has its own harness and also lets a team benchmark against other harnesses, including Codex. The output looks like published SweBench or TerminalBench results, run on the company's own work.

  On one task set, one model was 95% correct and cost six bucks on average, while another was 93% correct and cost two bucks. The models in the chart are slightly older because the field changes fast.

  The scores feed a model routing strategy or a harness routing strategy written

  ▶ at 16:49 Share Transcript

• What Is a Software Factory?

  Zach Lloyd defines a software factory as an agentic system in the cloud where a team defines the roles of various agents across building software for the team or for the product. The agents have a certain amount of autonomy and can do things on their own. The system should be very measurable and very closed loop. Over time the team can tune, measure, and benchmark it, and get confidence that people switching to agentic development are doing it in the most scalable and efficient way.

  When he talks to companies or developers, most of the world is trying to build some sort of internal software factory or a set of software factories. People trying to build their own often start by stringing tog

  ▶ at 2:23 Share Transcript