How Many Companies are Actually Using a Software Factory?

Probably thousands of companies, rather than millions. My working estimate is about 4,000, with a deliberately wide scenario range of 100 to 45,000 worldwide. That is an estimate built from disclosed organization counts and explicit assumptions, not a measured market size. The public evidence cannot yet tell us the real number.

What it can tell us is more interesting than another adoption headline: agent-written code has become ordinary much faster than agent-run delivery.

GitHub reports that agents supplied the first draft of more than 40% of pull requests in September 2026, up from 17% in March. That is more than a 2.35-fold increase in share in six months. It is platform telemetry, reported by GitHub, and includes private repositories. It counts a detectable starting point, not who reviewed, corrected or merged the work. (Octoverse, October 28, updated November 6)

A first draft and a factory are different units of progress.

What counts as a factory?

For this count, a software factory is a recurring system in which coordinated agents take real work from an issue through code, independent review, corrections and merge, with little human typing between stages. A person still sets direction, makes consequential decisions and can authorize deployment. The system does not have to deploy without permission to qualify.

That gives us three nested tiers:

Tier	What the company actually does
AI assistance	Someone uses completion, chat or an assistant for software work.
Autonomous coding	An agent takes a task, edits code, runs checks and returns a pull request. A person may coordinate everything after that.
Software factory	Agents coordinate the recurring issue, coding, independent review, revision and merge loop.

A multi-agent harness can run several agents at once without doing the third thing. A multi-agent workflow can orchestrate a review or a CI task without owning the whole loop. Using several models is not, by itself, either one.

I count a company once if a team there runs the workflow. I include solo businesses and exclude one-off demos. This is participation, not a claim that an entire enterprise has changed how it develops software.

The gap shows up in the pipeline

JetBrains’ 2026 CI/CD report supplies the most useful contrast I found. Roughly half of the surveyed organizations use AI for some CI/CD-related task. Only 8% report an AI step running inside a build or test. Another 28% reach their CI/CD tool through MCP or agent skills.

The main CI/CD survey had 598 respondents; the report also uses separate surveys, which it does not pool. These are vendor-research results, not a global company census. An AI build step can be much less than a factory, and a factory can run outside CI. Still, the distinction is revealing: helping around a delivery system is further along than becoming the delivery system. (JetBrains, 2026 report)

The broad developer surveys measure something broader still. Stack Overflow’s 2026 coding-tool question combines assistants and agents and asks about past-year use. Its methodology records 30,903 valid responses, recruited from its community, while individual question bases are smaller. JetBrains reports 90% weekly coding-agent use among its professional-developer sample from May–July. Neither asks how many companies run the issue-to-merge loop above. (Stack Overflow methodology, AI questions, JetBrains adoption study)

The vendors are selling a transition

The commercial momentum is real. It just measures several different things.

Disclosure	What it establishes
Microsoft, April 29, 2026: nearly 140,000 organizations use GitHub Copilot.	A large product footprint, including assistants and agents. Organizations are not deduplicated companies. (Earnings call)
Cursor, February 9, 2026: more than 50,000 teams.	Team adoption across Cursor, not background agents alone. (Cursor)
OpenAI, June 2, 2026: over five million weekly Codex users, above six times February’s figure.	Rapid user growth across coding and other work, not a company or cloud-factory count. (OpenAI)
Anthropic, February 12, 2026: Claude Code above $2.5 billion in run-rate revenue; business subscriptions quadrupled since January.	Spending and subscription growth, not the number of autonomous pipelines. (Anthropic)
Cognition, May 27, 2026: $492 million run-rate revenue and enterprise usage above ten times the start of the year.	A growing agent business. Usage is not unique customers. (Cognition)

All those figures are self-reported, and their dates matter. Revenue run rates are annualized rates, not revenue already earned over a year. Adding users, teams and organizations together would count some companies repeatedly.

Factory makes the strongest explicit claim about the final tier. In September it named Nvidia, Blackstone, RBC, Palo Alto Networks, Adobe and T-Mobile as companies building factories on its platform. It did not publish a total factory count. Its November acquisition of HumanLayer adds another overlap problem, not another population we can simply add. (Factory, September 15, HumanLayer acquisition, November 4)

Product availability also moves faster than adoption statistics. OpenAI introduced dots on September 29 without a separate customer count. Google introduced Antigravity 2.0 in May, but made Jules read-only after November 11 ahead of its December shutdown. Counting product launches as continuing installations would get even the direction wrong in that case. (Dots, Antigravity, Jules changelog)

We can see the work, but not all the companies

There is a public dataset. AIDev contains 932,791 agent-attributed pull requests across 116,211 repositories. The paper appeared in February 2026, but its data stop on August 1, 2025. Those repositories are not that many businesses, and neither private factories nor unsigned agent work are fully visible. (AIDev)

A study of its popular-repository subset found overlapping agent PRs in 40.2% of repositories. Only 0.5% of overlapping pairs involved different tools. That does not mean multi-agent work is absent: one tool can run many agents. Nor does overlapping work prove the agents coordinated with each other. (Xu and colleagues, July 2026)

An actual census would follow recurring work through every stage, record human interventions, deduplicate company identities and include private repositories with consent. A bot account or a co-author trailer alone cannot answer those questions.

My estimate, with the missing pieces exposed

I used Microsoft’s April organization count as an anchor, rather than inventing a global number of software companies. Then I varied the conversion to companies, the coverage of other vendors after overlap, and growth to November. I estimated what fraction delegate recurring coding tasks, and what fraction of those operate the stricter factory loop.

Companies with recurring use, November 2026	Low scenario	Central scenario	High scenario
Any AI software assistance	About 200,000	About 700,000	About 2 million
Autonomous coding agents	About 20,000	About 180,000	About 900,000
Software factories	About 100	About 4,000	About 45,000

These are judgment scenarios, not measured bounds or a confidence interval. The central factory case is a working guess, not a statistically most likely answer. The true count could lie outside the range.

Here is the central arithmetic: 140,000 organizations × 0.85 companies per organization × 4 for all-vendor coverage × 1.5 for subsequent growth = 714,000 companies using assistance. Assume a quarter delegate recurring tasks, and 2% of that group operate factories: 714,000 × 0.25 × 0.02 = 3,570.

Everything after the disclosed anchor is an assumption. In particular, no public measurement establishes that last 2%. Change it to 0.5% or 5%, leaving everything else fixed, and the result becomes roughly 900 or 9,000. That is why I would describe this market as thousands, with low confidence, rather than pretend that the central number has been counted. The full source ledger and low/high arithmetic are committed with this post’s research notes.

How fast is it growing?

Agent adoption is growing fast. A constant-definition factory-company growth rate is unavailable.

For multi-agent harnesses, the June Codex usage study found more than 10% of users managing at least three concurrent agents during a week. That measures parallelism, not an end-to-end factory. OpenHands had 64,000-plus stars in its November 2025 SDK paper and 93,522 at my November 15, 2026 GitHub API check. That suggests less than about 46% star growth, allowing for the rounded old count. Stars are interest, not deployed companies. (Codex research, OpenHands SDK paper, GitHub API)

For workflows, GitHub moved Agentic Workflows from technical preview in February to public preview in June. A2A had more than 150 supporting organizations at its first anniversary in April. Those establish available infrastructure and institutional support. Neither supplies a factory-company growth rate. (GitHub, Linux Foundation)

There is growth in the plumbing: npm counted 14.74 million LangGraph package downloads in September and 16.83 million in October, about 14% more, or 10% more per day after allowing for the longer month. Those are registry requests, including repeated CI installs and noncoding uses, not new factories. (September counter, October counter, checked November 15)

Analyst forecasts do not fill the gap. Gartner’s July forecast concerns smaller engineering teams by 2029, not factories operating today. Forrester’s June assessment distinguishes broad agent adoption from the rarer production multiagent system. DORA’s checked annual-research archive still links 2025 as its latest report; I found no 2026 factory census there. (Gartner, Forrester, DORA)

What I would do now

I have one documented example of the stricter workflow: the agent teams I directed merged 141 pull requests in one day. That illustrates what I mean by a factory. It does not tell us how many other companies have one, or what output another team should expect.

For a team deciding now, the useful experiment is to take one recurring class of work through the entire loop. Give coding and review separate owners. Define tests, merge rules and escalation paths. Measure completed changes, rework, defects and human interventions. More simultaneous agents without a reliable handoff can simply produce a larger review queue.

I am building Astrohacker for this transition. The opportunity is to make the complete loop repeatable and trustworthy. Agents already produce plenty of drafts. The harder work is building the system that can finish them.