The Customer Is the Builder: How AI Coding Agents Are Rewriting the Economics of Software
B · Useful, limited University of Utah researchers assemble a careful, reproducible set of public indicators and flag their own caveats, but the core evidence is extrapolation plus a third-party survey of self-reported decisions.
Summary based on the full text·Original: about 60 minutes to read
In short
A preprint (arXiv, 21 October 2026) by Ripon Chandra Malo, Zimeng Yu and Tong Qiu of the University of Utah. It asks whether coding agents let organisations build software they used to buy. The authors define a "software substitution frontier" (functions a customer could build and maintain for less than buying them) and line up leading indicators of building capacity against lagging indicators of purchasing. Capacity is rising fast; purchasing has barely moved in aggregate, though some individual buying decisions are changing.
Key numbers
- about every 129 daysdoubling time of benchmark task length leading agents complete at 50% success, since 2023 (95% interval 102 to 154)
- 575potential open-source alternatives found for selected commercial products (from 1,211 candidates)
- 32 percentrespondents to a McKinsey organisational survey whose organisation skipped a software purchase to build with coding agents
- 18.6 percentStack Overflow 2026 respondents who build in-house and said they were replacing paid vendors
As reported by the source.
English·Web page·opens arxiv.org
What it reports
- Agent task length at 50% success has doubled about every 129 days since 2023 (METR data).
- A broad search for twelve commercial products found 1,211 candidates, of which 575 are potential open-source alternatives; some matches still need validation.
- A public case (the Craft suite) produced 12 application repositories in 12.0 days of development, as early versions of unverified completeness.
- Market evidence is mixed: purchased software gained US investment share, and in-house built software fell from 14.0% of software investment in 2022 to 13.3% in 2025.
- McKinsey's survey (1,719 responses, 97 countries, May-June 2026): 32% said their organisation decided against buying at least one product or feature because it could build it with coding agents.
- Stack Overflow 2026: of 13,338 respondents who build in-house, 77.3% build small scripts, 63.0% team tools, 18.6% say they replace paid vendors.
- Stack Overflow monthly questions fell 99.1% from the 2022 level, a decline that began before ChatGPT.
How they got it
The authors split the argument into links that can fail separately (agent capability, software creation, competition with paid products, customer switching) and assemble public series for each: METR task lengths, GitHub repository counts, package-registry additions, BEA software investment, Homebrew install analytics, developer and organisational surveys. They fit capability trends and propose conditional projections and market tests rather than forecasts.
Limits
- The authors state the projections are extrapolation, not a forecast of switching.
- The 32% figure is self-reported, from a survey the authors did not run; it does not measure spending displaced or tool quality.
- Open-source alternatives and rapid reimplementations show creation, not reliability, security or maintenance.
- Developer survey respondents are self-selected.
Why it matters for companies
Software buyers, including HR and finance functions with large SaaS estates, now have a credible build option for narrow tools, which changes negotiating position even when no subscription is cancelled. The paper's own caveat is the useful discipline: the cost of building is falling, the cost of maintaining, securing and integrating is not shown to be. Build-versus-buy reviews should price ongoing ownership, not the first version an agent produces in a week.
Summary based on the full text. Read the original for the full argument and data.Read the original on arxiv.org
- Tags
- Coding agents