Cisco’s Unlimited-Token Refactor Shifts the Bottleneck From Writing Code to Proving It
Cisco president Jeetu Patel has given one engineer unlimited AI tokens and 24 months to refactor 98% of the company’s legacy code. The cost ceiling that normally stops a project of this size is gone, which leaves a harder question: how Cisco will prove the refactored code still behaves like the original.
A Four-Week Proof of Concept Comes First
Patel described the plan to Fortune at a small roundtable dinner, and the outlet published the account on October 30, 2026. The engineer is Padmanabhan “Paddy” Krishnamurthy, 27, who cofounded the real-time translation startup EzDubs in 2023 and joined Cisco when it acquired the company in 2025. Patel told Fortune he found Krishnamurthy five levels below him in the organization, working on something Patel called unstrategic.
Krishnamurthy told Patel he could rewrite Cisco’s legacy code with a team of two, and Patel said even the most AI-forward people on his own team were not thinking that way. A four-week proof of concept comes first, with the 98% target due within 24 months. On tokens, Patel’s instruction was to “go nuts.” Fortune’s report does not say which products or how many lines of code count as legacy, and it carries no results from the proof of concept.
The coverage uses “refactor” and “rewrite” interchangeably, and they are not the same job. Refactoring restructures code while keeping its behavior identical; a rewrite can change what the software does. Patel’s own word was refactor, which sets the stricter standard: whatever the AI produces has to act exactly like what it replaced.
Generation Gets Cheap, Verification Does Not
Earlier large migrations show where the effort goes once generation costs little. Airbnb moved nearly 3,500 React test files from Enzyme to React Testing Library in about six weeks, against a manual estimate of 1.5 years. Its pipeline ran each file through automated Jest, lint and TypeScript checks and retried failures up to a set limit. A first bulk run reached 75% of files in four hours, and four days of prompt and script refinement lifted completion to 97%. Engineers fixed the remaining 100 or so files by hand, which took about another week.
Google reported a similar pattern in a paper on its internal migrations. In an int32-to-int64 ID change across the Google Ads codebase, which spans more than 500 million lines, AI wrote 80% of the code modifications in landed changes, and engineers estimated a 50% cut in total time. The authors also reported that in most cases the engineer had to revert at least some of the model’s changes, and they named the speed of human review as the bottleneck.
GitHub’s numbers show where that load lands. The platform recorded 7.38 billion commits in September 2026, more than five times the figure a year earlier, and GitHub Actions, its workflow automation service, ran 3.26 billion times, more than four times as often as a year before. The implication is that every company letting agents write code at that pace inherits the same multiplier on its own build and test systems.
Cisco Has Already Named the Standard
In an interview with SDxCentral, Patel said organizations should prepare for a world in which humans cannot review all the code AI writes, and that teams must “automate the test beds and harnesses” that measure outputs. He added that whoever evaluates those outputs should answer for the code, and that every piece of it should trace back to a human. In the same interview he said Cisco’s AI Defense product was built entirely with AI, and he set targets of about half a dozen fully AI-written products by the end of 2026 and roughly 70% of the portfolio by the end of 2027. Those are targets, not results.
One way to read the legacy refactor is as the older-code counterpart to that strategy, facing the same test. The implication is that Cisco’s test coverage, more than its token spend, will decide whether the 98% figure holds. Fortune’s report does not say how Cisco will measure it, and the four-week proof of concept is the first chance to show a method exists.
What to Have in Place Before You Try This
Fortune’s report does not describe Cisco’s method, so what follows is advice drawn from the Airbnb and Google cases and from Patel’s stated standard, not a description of what Cisco does. If you plan a large AI-driven refactor, settle three things before you spend the first token.
- Test coverage that can fail the code. Airbnb’s pipeline worked because every migrated file faced an automatic judge: it passed the Jest, lint and TypeScript checks or it went back for another attempt. Production code needs the same kind of referee. Map which of the modules you plan to change have tests that would catch a behavior change, and write tests that record what the code does today for the ones that do not. Do that before the model touches them, because a test written against the model’s output can only confirm what the model already did.
- A named owner for every change. Patel’s stated standard is that whoever evaluates AI output should answer for the code, and that every piece of it should trace to a human. Google’s paper found that engineers reverted at least some of the model’s changes in most cases, so someone has to read each diff and sign for it. Assign owners by module before the run starts. Then budget their review hours as a line item, since Google named the speed of human review as the bottleneck.
- A pass-rate threshold fixed in advance. Decide before the first run what share of changes must pass your checks unchanged for the project to continue, and what happens to the rest. Airbnb’s first bulk run cleared 75% of files in four hours, four days of tuning lifted completion to 97%, and the final 100 or so files took about another week by hand. The long tail is where the schedule goes, so staff for it. The right number depends on what a failure costs in that system: a test-file migration can tolerate more misses than a payments service.
Cisco’s four-week proof of concept is where these three items will show up first. If Cisco discloses how much of the refactored code passed its existing tests unchanged and how many reviewer hours the rest consumed, the industry gets rare evidence from a company that builds networking infrastructure. Without those numbers, the 98% target stays an ambition with an open-ended token allowance behind it.
The post Cisco Gives a Former Startup Founder an AI Budget With No Ceiling appeared first on DataFLOQ.
