A software factory, as its sellers use the phrase, is a fleet of coding agents wired across the whole delivery loop - triage, spec, implement, review, verify - and sold as a plant rather than a tool: something you configure once, run continuously and judge by its output. The word does positioning work a “coding agent” cannot. An agent is bought per seat by a developer; a factory is bought by whoever owns throughput, and it arrives with a metric already attached.
The question this thread follows is whether that is a category or a company. As of 2026-11-10 it is a category with a price shape but no price, and a metric with many values but no definition.
- The word is everyone’s, and it is now a product name at five sellers.
- The price is a markup on the model maker’s list price, published at three sellers, from 10% to 40%. The highest belongs to the incumbent.
- The metric has figures from sellers, buyers and a measurement firm, and no two count the same thing.
The incumbent’s rate was on another domain
The product Harness sells as its Cosmos Software Factory Agent (opens in a new tab) has a public price. It is not on Harness’s pricing page (opens in a new tab), which on 2026-11-10 still had no Cosmos line and sold its plans “by quotation”. It is on Augment’s (opens in a new tab), the coding-agent maker whose Cosmos Harness took over on 2026-10-08. That page’s footer now reads ”© 2026 Harness, Inc.”, and its navigation calls Cosmos “Software Factory”. Augment’s product page (opens in a new tab) is headed “Build your software factory.” The terms on 2026-11-10:
- Model usage is “billed at the provider’s public API list price”.
- A service fee of “a flat 40% fee on LLM usage (no fee on compute)”.
- Compute is Cosmos time, at $0.19 an hour (opens in a new tab) in five-minute steps.
- Plans: Standard $20 and Business $100 a month, flat for up to 50 seats, with that much usage included. Enterprise is custom.
Put the three published markups side by side:
- SuperPlane: its tokens at the provider’s API rate plus 10% (opens in a new tab), under the heading “Pricing for your software factory” (1.0 launched 2026-10-27 (opens in a new tab)).
- Warp: pay as you go (opens in a new tab) at “20% above API rates”, Build at $20 a month “at API rates”, Business $50 a user.
- Augment, for Harness: list price plus 40%.
That is a convention, a markup on someone else’s model rate, with a spread of four to one. A reader pricing a product in this space has a reference shape. They do not yet have a reference number.
The other incumbent never needed a markup. GitLab’s governed software factory (opens in a new tab) release (2026-10-06) named credits without rates, but the rates were public all along. GitLab’s pricing page (opens in a new tab) sells credits at $1 each, with $12 a user a month included on Premium and $24 on Ultimate as a limited-time offer. Its docs (opens in a new tab) list how many requests a credit buys on each model, and price a code review at a quarter of a credit. The incumbents, in other words, both publish a rate. They publish it in different units.
The startups with the most money behind the word publish none:
- 8090, whose product is “Software Factory®” and which raised a $135M Series A, describes a services engagement: “Our engineers build with your team … We host it, maintain it”. Its pricing page (opens in a new tab) serves the same text and a “How is it priced?” question with no answer.
- Tessl launched Factory on 2026-11-06 (opens in a new tab) (“over $120 million” raised, per the release). Its pricing page (opens in a new tab) does not mention Factory and prices agent sessions in credits “scaled to each session”.
- Factory, the company, sells seats with rate limits (opens in a new tab) and states no rate.
Every direction has a figure, and none match
Cost per merged PR now has values from every kind of source. Each counts something different:
- Tessl says its cost per PR fell from $75 to $13 (opens in a new tab). The release gives no window, no count, and no word on whether the PRs were merged. Tessl has no launch post of its own. Its blog said on 10-28 (opens in a new tab) that a factory’s cost should include “infrastructure, retries, review, and rework”. The release does not say whether the $13 does.
- Guild, which also sells a product called Software Factory, says “around $12 of model spend per merged PR” (opens in a new tab) over “a recent quarter” (2026-09-11), with no count.
- Warp has two figures for its own team off one $80 baseline: $30 on 2 September (opens in a new tab), under a “75%” headline that $80 to $30 does not make, and “$80 → $20 in the last six weeks” (opens in a new tab) on 15 October. Its 7 core components (opens in a new tab) post (10-29) names cost per PR as a dashboard metric without defining it. No PR count stands behind either figure.
- CampusIQ, a buyer, published “$53,000 divided by 1,832 merged pull requests” (opens in a new tab), or $29. That is token spend only, over a window the post does not date.
- A smaller buyer: Mateusz Piwnicki’s playbook (opens in a new tab) for two Cursor cloud agents at an online shop counts 232 PRs merged of 259 opened between 26 September and 2 November, at “about $7” per merged PR as a two-week median. It does not say what the cost includes.
- Jellyfish, which sells engineering measurement, publishes the only cross-company figure, in its October 2026 report (opens in a new tab). It counts salary plus AI usage, divided by merged PRs opened by humans. The median company’s cost fell “from $1,202 to $820” over a year, and “AI’s direct slice of the cost is small, about $18 of the $799 average.”
All of it is self-reported, and none of it lines up. The sellers’ figures are tokens on agent PRs. Jellyfish’s is everything, on human PRs. Uber, which named cost per merged PR as the metric for its in-house Software Factory (opens in a new tab) on 2026-08-27, has still published no value.
The Jellyfish number is the most useful thing here for a reader, because of what it implies. If AI is about 2% of what a merged PR costs, a buyer who owns throughput will judge a factory on the other 98%. The gap between a 10% and a 40% markup is a gap of about one percent of that bill.
The word is everyone’s
Since late August the word has spread to every kind of company in the market:
- A buyer. Uber.
- Two incumbents. GitLab (2026-10-06) and Harness, through Cosmos (2026-10-08), which Augment had already been calling a software factory.
- Product names. 8090’s Software Factory®, Tessl’s Factory, Guild’s Software Factory, Tessl’s agent (opens in a new tab) page, and SuperPlane’s “open source software factory”.
- A relaunch. Opsera relaunched as “the context-aware software factory” (opens in a new tab), per SD Times (opens in a new tab) on 2026-10-29.
- The company named Factory. Its Typedef acquisition (opens in a new tab) (2026-10-30) is headlined “Bringing Data Context to the Software Factory”.
- The explainers. Faros (opens in a new tab), LaunchDarkly (opens in a new tab) and Port’s research index (opens in a new tab), each selling something that sits inside the factory.
The word was never Warp’s to begin with. The US Department of Defense has called its in-house units software factories (opens in a new tab) for years. For a reader positioning a product, “factory” no longer sets it apart.
The tension
The two sellers with the most money behind them still sell other nouns. OpenAI’s frame is Codex as a platform (opens in a new tab). No title on Cognition’s blog (opens in a new tab) since September uses the word, through “Introducing SWE-2 Flash” on 11-04.
The best arguments against the frame are about what it measures. PostHog’s newsletter notes that “there’s no actual criteria for what counts as a software factory” (opens in a new tab). Noah Brier argues that the hard part is keeping agents’ work coherent with what the company is trying to be: “It’s a software company, not a software factory” (opens in a new tab). And Jellyfish’s $18 in $799 says the unit sellers compete on is the smallest line of the bill.
So, as of 2026-11-10:
- “Software factory” is everyone’s word and, at five sellers, a product name.
- The price has a shape, a markup on the model maker’s list price, and three published points: 10%, 20% and 40%. Both incumbents publish a rate. The best-funded startups do not.
- Cost per merged PR has six published values and no shared definition.
The category exists. Whether it has a price depends on whether the markup settles on a number, and whether a buyer can ever compare two factories depends on two figures being counted the same way.
What would settle it
3 open loops
-
01 By end of March 2027
Does the markup settle on a number? Three sellers publish one: 10% (SuperPlane), 20% (Warp's pay-as-you-go) and 40% (Augment, for Harness's Cosmos). Harness's own pricing page still quotes, 8090 serves no price, Factory sells seats with limits and Tessl prices sessions in credits with no rate. A fourth seller publishing inside that band, or Harness putting Cosmos on harness.io at a stated fee, makes the markup the category's price. A seller dropping its public rate for quotes points the other way.
-
02 By end of January 2027
Do any two published costs per merged PR count the same thing - merged PRs, a dated window, and the spend named? CampusIQ's $29 is tokens over an undated window; Tessl's $13 has no window, count or definition; Guild's $12 has no count; Jellyfish's $820 is all-in and counts only human-opened PRs. Uber defines the metric and still publishes no value.
-
03 By end of 2026
Does OpenAI or Cognition adopt the word in its own launch copy? GitLab, Harness, Tessl, Guild, Opsera and Factory have. Through 2026-11-10 OpenAI still sells a platform and Cognition sells Devin; no title of Cognition's posts through "Introducing SWE-2 Flash" (11-04) uses the word. Adoption by either makes "factory" the field's name rather than everyone else's frame.
Each is checkable by a date. If one goes quiet, that is itself an answer.