Zach Lloyd opens Warp’s internal agent dashboard and follows a mobile-app task that began in Slack. The record keeps the conversation and computer-use screenshots together, so the team can inspect how work happened.

In this Latent Space interview, Warp’s founder and CEO argues that teams should manage coding agents as shared infrastructure. His proposed “software factory” has a versioned workflow, a record of its work and tests for improving it. The November 3, 2026 recording runs 34:48.

Put the workflow in code

At 04:30, Lloyd describes a factory for a specific product surface, such as a mobile application. Its definition names repositories, skills, integrations and steps such as triage, specification and implementation.

Warp uses YAML for that definition. The benefit is ordinary source control: a team can review changes, branch experiments and roll back a bad configuration. Coding agents can also propose edits to the process they run.

Product engineers submit work through familiar tools. Slack, Linear, Jira and GitHub become entrances to the same workflow. Platform and developer-experience teams inspect the shared record and tune the machinery behind it.

Count what the agents cost, including inspection

The measurement tour covers shipping speed, throughput, human interventions per merged pull request and cost. Lloyd wants teams to break spending down by person, automation and development stage.

Warp’s “scorers” use a judge model, an AI model grading another model’s work, to inspect stored runs. One example checks for redundant tests left behind during implementation. Teams choose the grading prompt and sampling rate. That last setting matters because inspection itself costs money.

This gives teams a way to find recurring problems. Whether a scorer’s judgments track software quality remains a separate question. Lloyd acknowledges there is more work to do on quality measurement.

A proposed improvement still needs a test

At 14:32, Lloyd shows an improvement agent identifying 54 triage failures. It proposes allowing the triage agent to spawn a subagent for investigation. The suggested change becomes a pull request against the factory definition.

At 15:41, he states the missing step: automatically replaying those failures to check whether the proposed change would actually help. The demonstration shows agents proposing process changes. Validating each proposal against the affected runs is described as future work.

A separate benchmark view compares configurations on selected historical tasks. Teams can vary the model and the harness, the software that runs agents’ tools and commands. Lloyd wants those results to guide routing decisions for the team’s own codebase.

Keep the parts replaceable

From 19:23, Lloyd sketches the broader stack: workflow definitions, stored evidence, integrations, permissions, compute, inference, orchestration and improvement tools.

He says Warp primarily uses Namespace for execution environments. His architectural recommendation is to make that layer replaceable. Different tasks also need different machines: testing an iOS application may require a Mac virtual machine and simulator.

The same flexibility applies to inference providers and agent software. Customers may already have provider commitments they want to use. Lloyd presents the factory as infrastructure that coordinates these choices. His blueprint is a proposed approach, not a ratified industry standard.

The reported savings belong to Warp’s workload

At 23:59, Lloyd says internal tuning reduced Warp’s cost per pull request from about $80 to roughly $15–20. He cites expensive triage models and excessive subagents as earlier sources of waste. He also reports 40% higher throughput over the previous month, without a higher revert rate.

Caveat: these are vendor-reported internal results. The interview supplies neither an independent comparison nor a full accounting of review and platform costs.

Lloyd also limits the audience. A few-person team should generally keep using existing coding subscriptions. He suggests five to ten engineers as a possible starting point for shared infrastructure, with fit depending on the team’s systems.

The closing distinction is between product engineers and factory engineers. One builds features. The other improves the workflow that builds them. Lloyd’s case depends on preserving enough evidence to tell whether those improvements work.

Source: Latent Space, published November 3, 2026, 20:16:25 UTC. Full caption coverage: introduction through closing at 34:46.

Listen to the podcast audio.