The field in July 2026
Everything on this page moves. The chapters are written to still read correctly in a year; this page is not. It records who was doing what in July 2026, with each figure carrying the status of its evidence, so that a reader arriving later can tell what to distrust first.
The short version: light factories are the production reality, and the one well-documented dark factory is expensive.
Who runs what
| Organisation | Level | Volume | Who reads the diff |
|---|---|---|---|
| Stripe — “Minions” | L3–L4 | 1,300+ agent PRs merged per weekprimary | human, all of them |
| Spotify — Honk / Fleet | L3–L4 | ~1,000 PRs every 10 daysreputable | human, CI-gated |
| Cursor — cloud agents | L3–L4 | 35% of internal merged PRsprimary | human — the stated bottleneck |
| StrongDM — Software Factory | L5 | not published | nobody, by rule |
Stripe: the light factory at scale
Stripe’s February 2026 engineering deep-dive describes Minions merging over 1,300 pull requests per week carrying zero human-written code, all of them reviewedprimary. The pipeline is a five-layer blueprint mixing deterministic and agentic nodes; work arrives from Slack, bug reports and feature requests; agents run in isolated cloud devboxes against 3M tests and 500 MCP tools; and a two-round CI limit bails out to humans on complex debuggingprimary.
The write-up’s own caveat is the most useful part for anyone planning to copy it: this was built on a decade of prior investment in standardised devboxes, test suites and internal tooling. The agents are the last layer, not the first.
Spotify: migrations at fleet scale
Presented at QCon London in March 2026 and reported by InfoQ: Honk merged 1,500+ PRs in its first nine months and now merges around 1,000 every ten days (work that took roughly three months half a year earlier), with 60 to 90% time savings on migrationsreputable.
Two details are worth more than the headline. Deterministic scripts still handle about 70% of migration volume; Honk targets the remaining 30% of edge cases. And the LLM judge, which initially vetoed around 25% of sessions, was removed by March 2026 as models improvedreputable: a rare documented instance of a gate being retired on evidence rather than on hope.
Cursor
35% of internally merged pull requests now come from cloud agents operating in sandboxes; the February 2026 computer-use launch cited more than 30%primary. Agents build, test through the browser UI, record video proof and produce merge-ready PRs. Code review remains the human bottleneck, and the company’s “self-driving codebases” language is explicitly aspirational.
StrongDM: the reference dark factory
Founded July 2025, publicly described February 2026. Two rules: code must not be written by humans, and code must not be reviewed by humansprimary. Its machinery is covered in chapter 8: scenario-as-holdout, the Digital Twin Universe, “satisfaction” as a probabilistic metric.
The economics are the part most often left out of the retelling: a stated benchmark of at least $1,000 per day per engineer in tokens, roughly $20,000 a monthprimary.
Consultancies and the enterprise middle
Thoughtworks launched AI/works in January 2026, an agentic delivery platform behind its “3-3-3” model of idea to production in 90 daysprimary, while its Technology Radar has been the loudest practitioner voice urging restraint, warning about cognitive debt, semantic diffusion, AI-accelerated shadow IT and complacency with AI-generated codeprimary. Accenture launched a forward-deployed engineering programme with ServiceNow in May 2026reputable. EPAM showcased Claude Code-based SDLC automation citing expected gains of ~35% in development and ~50% in supportvendor-reported. Those are projections, not measurements. Cognizant partnered with Cognition in January 2026reputable.
The reality check
Nearly all production deployments are Level 3 to 4. True Level-5 operation is confined to StrongDM and, by Shapiro’s own estimate, a handful of small teamsprimary.
And the capability numbers behind the deployments are softer than the leaderboards suggest. IBM Research found frontier agent capability dropping from about 73% to about 11% on discriminative subsets, with novel-issue resolution around 18 to 20%primary.
What is missing here
This page reports what the sources record. It does not include: any organisation running dark without publishing it, which is likely several; the non-Western market, which the source material barely covers; and anything from firms that tried and quietly stopped, which is the category most likely to be informative and least likely to be written down.