Nobody built it, so nobody can fix it

Meta ran the experiment at full scale, then pulled the plug

On August 25, McKinsey told everyone the water was fine: 27% of enterprises have decided against buying software because agentic coding tools can build the thing in-house. The next morning, Reuters published what the water actually did to the one company that swam furthest out.

Meta called it Project OT. Sketched at a January leadership retreat at Zuckerberg's Hawaii estate: an "AI native" company, agents doing the daily work, some teams cut by as much as 60% in two waves, humans regrouped into small pods of general-purpose "builders." Then Meta looked at its own dashboards. Code changes to its AI platforms were up 220% year over year. New features actually reaching users: up 36%. Technical and security incidents: up 40%, with employee time spent firefighting them up 70%.

Zuckerberg killed the second wave. His own words: "The trajectory of the agentic development over at least the last four months hasn't really accelerated in the way that we expected."

Meta printed code. It could not print understanding. Triple the code, a third more product, and a fire alarm that never stops ringing. That is the full-scale result, from the company with the deepest engineering bench on the planet and every incentive to make the story end differently.

The 27% just bought Meta's problem without Meta's bench

Now go back to McKinsey's 27%, because the survey hides the interesting part. Among firms over $1B in revenue, agent programs running at scale jumped from 27% to 40%. Enterprise-level profit impact from AI: stuck at 37%, the exact same figure as last year. McKinsey's own verdict: "Organizations' conviction in AI is growing faster than the immediate financial returns they can attribute to it."

The 27% read their agent demo as a free lunch. Skip the license, skip the vendor, let the agents build the approvals tool over a weekend. And on day one they are right. The tool works, the invoice dies, the CFO gets a slide.

Here is what I think day ninety looks like, and this part is opinion, so weigh it against my vantage as an operator who builds agentic systems for a living. Meta had thousands of engineers who could open the hood when the 40% incident spike hit, and it still retreated. The mid-market company that deleted the license AND never had the bench has no hood-openers at all. It has a pile of code that passed the demo, written by a tool, prompted by a manager, reviewed by nobody.

The tools ship broken product, and nobody can see where

This is the part the build-it-yourself crowd has not priced. Most AI coding tools ship broken product, and the breakage is invisible until production finds it for you.

Lightrun surveyed 200 senior SRE and DevOps leaders at enterprises with 1,500+ employees: 43% of AI-generated code changes need debugging in production AFTER passing QA and staging. Developers now burn 38% of their week, roughly two days, debugging AI-written code. And when a fix ships, 88% of teams need two to three redeploy cycles just to confirm it worked.

CloudBees found the same shape from the other side: 92% of enterprise tech leaders were confident their code was production-ready. 81% then watched production issues rise with AI code adoption. 69% cite security vulnerabilities introduced by AI-generated code.

Call it nobody's code: software nobody wrote, breaking in ways nobody can find. When a human builds a system, the debugging map lives in a person. You walk to their desk, they say "check the queue handler, I always hated that part." When an agent builds it and the builder-of-record is a prompt, that map does not exist. In Lightrun's survey, 74% of financial services teams fall back on tribal knowledge to diagnose incidents. Tribal knowledge of code no tribe wrote.

The loneliest thing in software used to be a stack trace in a codebase whose author left the company. At least the author existed.

The cleanup crew has already been hired

If you want to know how this movie ends, watch what the sellers are buying. On August 13, OpenAI signed IBM: a dedicated OpenAI Practice, thousands of consultants and engineers certified to deploy GPT-5.6, Codex and ChatGPT Work inside enterprises. IBM Consulting's Andy Baldwin, in the press release: "The challenge is not access to AI technologies." The challenge, he says, is integrating AI securely, at scale, into complex enterprise environments.

Read that as a confession. OpenAI can see, better than anyone alive, what its own tools ship into companies. It just bought thousands of human engineers to stand between its models and production. The deployment army I wrote about in June is real, and it is exactly the bench the 27% think they no longer need. The bill on AI coding was always coming due; Meta just paid it in public.

The agents are staying. The orphan code doesn't have to.

The obvious counter: agentic coding is real, and I should know, my company ships agentic systems. The 27% did not hallucinate the savings, and some of them, the ones with real engineering depth, will build things a vendor never could. True. Concede it fully.

But the dividing line in the next two years will run somewhere very different from build versus buy. It runs between software somebody on your payroll understands under the hood, and software nobody does. Deployed and deployable beats generated and abandoned, every time it matters, which is always at 3am.

So before you skip the renewal, ask one question: when this thing breaks, who in my building opens the hood? If the answer is a name, build it, own it, enjoy it. If the answer is a vendor, keep renting without shame; Meta scurried back and so can you. And if the answer is nobody, you have not cut a cost. You have adopted an orphan. Nobody built it. Nobody will fix it.