Software teams are building factories for agents, not just faster tools for coders

The factory, not just the tool

Warp's Zach Lloyd describes a software factory as an agentic system in the cloud where roles and autonomy are defined for building software. He presents factory engineering as a new discipline, distinct from product engineering, focused on building the system that builds software. m:…8889

The factory itself should be defined declaratively in code, Lloyd argues, so it can be versioned and improved automatically. Centralizing agent conversations and artifacts is essential to a closed-loop system that improves over time, and self-improvement agents can even suggest pull requests that modify the factory definition code itself. m:…8889

This suggests the job has changed shape: the valuable work is no longer picking a coding model, but designing the system around it — how agents get tasks, where their work is checked, and how the system learns from each run. That framing is Lloyd's, not a measured industry standard, and it comes from a company selling factory tooling, so it should be read as a vendor's argument as much as a description. m:…8889

Why the plumbing broke first

OpenAI ran into the factory problem in its own infrastructure. The company replaced its monorepo's Git backend because Codex-driven activity caused reliability to fall below 90%. Its replacement, Pando, cut median push latency by nearly 5x, handles nearly a petabyte of Git data per day, and lets branches move independently rather than enforcing repository-wide consistency. OpenAI has not published code or APIs for Pando. m:…6943

Qodo, the company writing about OpenAI's move, argues that code review is the next bottleneck for agent-scale development. Its evidence is the volume shift: agent-started pull requests on GitHub increased from 17% in March 2026 to over 40% in September 2026, and about 87% of agent-started pull requests changed multiple files, compared with 79% for non-agent pull requests. m:…6943

The review burden lands on people: 47.7% of engineering leaders identify reviewing and validating AI-generated code at scale as their biggest quality gap. Qodo's own answer is a triage product that ranks open pull requests across Git providers by priority, blast radius and service-level agreement — a reminder that every diagnosis in this space arrives with a product attached. m:…6943

The execution layer gets funded

Daytona says it pivoted from automated development environments for humans to an Agent Execution Cloud, on the argument that AI agents require isolated, fast, stateful machines available through an API. The company says it grew from a $1M run rate to a $100M run rate in four quarters, a 37.5x increase in 2026, and offers instantaneous startup times, persistent state and per-second billing. It also says it rebuilt its product with only a few months of runway left. m:…0544

Daytona cites ICONIQ research showing software companies historically took five or more years to reach $100M in annual recurring revenue, while AI-native companies are reaching it in one to three years, according to the 2026 State of Scaling report. Its examples include Anthropic reaching $100M in annualized revenue in about a year in 2023, Cursor in roughly 12 months, Mercor 11 months after $1M, and Legora and Wiz in about 18 months. m:…0544

The pattern here is worth noting for a founder: the companies cited as fast growers are not just selling a model, they are selling the infrastructure around agents — sandboxes, execution and operations. Warp makes a similar cost argument, saying it reduced its cost per pull request from $80 to $15-$20 through internal tuning and factory engineering, and that individual $200-per-month agent subscriptions run 5 to 50 times more expensive than API-based factory usage. Warp also cautions that factory infrastructure is generally overkill for teams with fewer than 5 to 10 engineers. m:…8889

The outer loop closes

Traversal, which says it raised $120 million in Series B funding at a $1 billion valuation in a round co-led by Kleiner Perkins, Redpoint Ventures and Khosla Ventures, is building on the other end of the loop: production. Its Workers are proactive, long-running agents that work alongside engineering teams to gather context and identify root causes, and the company says monthly investigations on its platform grew from 513 to nearly 73,000 in 2026. m:…5486

The company's stated aim is self-driving production, where software systems autonomously detect, understand and correct issues. It introduced a Production World Model and Causal Search Engine in March 2026, and says its Traversal MCP feeds production intelligence back into development to help coding agents ship more resilient code. Roughly two-thirds of usage extends beyond root cause analysis into areas like alert triage and security. m:…5486

The demand side of that pitch is familiar to anyone running software: Traversal cites figures that 42% of developers now use AI to write a majority of their code, while 52% are spending more time debugging code once it hits production, and that unplanned outages cost Global 2000 companies $600 billion in 2026. This suggests the factory story is really one story told three ways — OpenAI rebuilding plumbing, Daytona selling execution, Traversal selling operations — and the open question is whether the factory discipline Lloyd describes becomes standard practice or stays a luxury for large teams. m:…5486