Zach Lloyd says most companies mistake heavy agent use for a software factory, which he defines as an agentic cloud system defined in code.
Warp cuts its cost per PR from ~$80 to $15-20 and lifts throughput 40% without more reverts.
Lloyd says factories are overkill for small teams and a poor fit for bespoke-infrastructure firms like Uber and Ramp.
Software factories turn scattered coding agents into measured systems
Zach Lloyd argues that agentic software development should be defined in code as a factory, one per product surface, with tracing, evals and cost metrics attached. The loop can then be measured, benchmarked and improved, and because the definition is code, agents can open pull requests against it as models change. Warp sells that infrastructure, betting the durable asset is the system that routes work to models.
Takeaways
A factory is one artifact per product surface. Lloyd treats a software factory as a concrete artifact defined in code, a YAML file listing repos and MCPs, one per product surface. Like Terraform, the definition is versioned and changes can be reverted.
Heavy agent use is not a factory. Lloyd says the common misconception is calling a pile of automations and heavy agent usage a software factory. A real one defines its agents' roles and grants a degree of autonomy inside a closed loop that can be measured and tuned.
Factories replace vibes with measured scores. Warp calls its online evals scores. Each picks a judge model, a judging prompt and a sampling rate, then retroactively grades stored runs, catching things like redundant tests agents leave behind. The audience is engineering managers and platform teams.
Warp cut cost per PR to $15-20. Cost per pull request fell from about $80, when the triage agent ran on Opus, to $15-20. Throughput rose 40% in a month with no rise in revert rate, and about 60% of Warp's work starts in Slack.
Benchmarks vary model and harness together. Warp replays curated runs across a matrix of configurations, testing its harness against Claude Code, Codex and Pi. On one internal task set GPT-5.6 SOL was 95% correct at $6 and GLM-5.3 Flash 93% at $2, which feeds routing.
Agents open PRs against the factory. In the self-improvement loop agents look for failures in the scores and open pull requests against the factory definition. Lloyd cites 54 triage failure cases where the triager should have spawned a sub-agent. He calls the discipline factory engineering.
Subscription economics block pure coding agents. A $200 subscription can carry $8,000 of API usage, so pay-as-you-go costs far more. Lloyd says Warp will not compete as a pure coding agent, since you cannot build one on inputs that expensive. The host counts Cognition and OpenCode as the independents left.
The factory stack
A data layer holding every agent conversation and artifact.
MCP plug-ins with different access for different parts of the organization, such as marketing and engineering.
Tap a name to go there · tap a line to see why they are linked
On this episode 1
Show 1
Pluggable compute, with Warp on Namespace and Daytona, Modal and E2B as alternatives, and Mac VMs for iOS testing.
Flexible inference, including burning down commitments with Anthropic, OpenAI or Google, plus ACP-compatible harnesses like Pi.