Xerox Machines of Loving Grace

How well large language models fare with the unknown is still an open question. What is not in question is that they are extraordinary pattern replicators — and since a huge share of their training data is code, the patterns they replicate best are codebases.^{1 1}On SWE-bench, Claude Opus 4.6 Thinking resolves 79.2% of real-world GitHub issues; on CodeElo, o3 sits at the 99.7th percentile on Codeforces.

Coding agents are incredibly good at pulling code out of thin air, and obstinately iterate until things work. However, they regularly start failing as the project scope increases. While this is generally blamed on the models not being smart enough, I think that most of the time the issue is actually underspecification. As discussed before, when working with coding agents testing is everything. Indeed, if your specifications can be very well formalised in terms of requirements to be matched and tests to be passed, you can consider your codebase already implemented.^{2 2}Modulo some hundreds of dollars’ worth of tokens.

Xerocoding 101 #

I recently experienced this very strongly myself, as I used Claude Code to migrate an hypertrophic Access codebase^{3 3}A company-wide ERP system built on 4 linked databases, containing 251 tables, 335 input screens and reports, and 37 custom code modules (around 10,000 lines of code). It handled operations for multiple subsidiaries across different countries, supporting several languages and currencies; and talked to external accounting and payroll software, featured FTP and emails integrations, and role-based access control and multi-tenant isolation. into a more modern MySQL+PHP+Laravel web app. Goes without saying: I know neither Visual Basic, nor PHP, nor Laravel. But with the help of an army of terminal windows, work that was estimated at a couple of years of legacy development, or several months of AI-assisted coding, was addressed in a matter of 3-4 days of uninterrupted agents runs (with some 3-4h of supervision per day). It was fast but not trivial: it required lots of tricks to organize agents to parse specifications from the old codebase; and translate them into the new one; and making sure no piece was lost along the way. This is currently not just a prompt and a button press, but given where we were one year ago, it is safe to estimate that we will get there soon.

So, we are already at a point where generating a whole production-ready project is mostly a matter of orchestrating and spawning the right number of agents and allocating some budget for token consumption;^{4 4}My implementation cost about $200 in tokens. soon enough, the tokens will be the only requirement, and replicating a whole software stack will be something everyone could afford. I like to call this specific usage of coding agents for cloning software xerocoding,^{5 5}Or photocoding? to distinguish it from vibecoding (where the direction and the objectives come from the user “vibes”).

The copyists at work #

Xerocoding is popping up more and more in the news. Coding agents have been by now used to rewrite entirely complex python packages to change their license from GNU to MIT, and allow commercial usage, a practice that has stirred quite some debate; to reimplement Next.js from scratch using end-to-end testing. A company that claims to be writing 100% of their code with AI^{6 6}My own pledge for 2026. is describing how they could build “behavioral clones” of existing software. In general, testing at the surface is trivially simple when the behavior of the codebase is already highly constrained by existing usable tests.

⊕ Code shaped by its tests. Needless to say, xerocoding will have huge impact on the world of software production. In February 2026, over $285 billion was wiped from software stocks. If replicating a software product costs a few hundred dollars in tokens, charging users a fixed or monthly fee for access to it becomes increasingly harder. What remains in the exclusive hands of who has been building software when the codebase loses value this way? Data that has been hoarded by some; complicated legal compliance, for applications that require it; and network winner-takes-all-users effects. But even where some providers maintain those strength points, code becomes dispensable.

What does software amount to? #

But beyond the market panic, there is a more interesting question awaiting: what does software “amount to”, ultimately? The moment an agent can fill in the dots of detailed enough requirements and testing surface, those become interchangeable with the codebase; writing code will be just compiling specs. Beyond the doomsday scenarios of xerocoding, a whole new world opens up. Xerocoding does not address the code that still has to be written (or better: the specs that still have to be written, in order to generate code).^{7 7}How to call the minimal specifications that can produce a codebase? Anticode? Code complement? (Edit: or a further suggestion i got: Hamel codebase.) Massive code generation can’t get too far in helping people untangle real-world data generation, organizational mess, or byzantine legal requirements.^{8 8}Those just keep growing with the amount of time new tools liberate… But more and more, people who have no familiarity with the nitty-gritty of code development will start dealing with software creation. We will probably need new languages, both for specifications and for the codebase implementation.^{9 9}Contrary to what I thought, LLM proficiency with a language does not scale purely with training data availability.

An exciting Cambrian explosion seems to be around the corner.