I have read a lot of AI hype headlines this year. This one stopped me mid-scroll. Cisco, worth around $429 billion, has handed a 27-year-old engineer an unlimited budget for AI tokens and told him to rewrite 98% of its legacy code within 24 months. Not 10%. Not a pilot project on one small service. Ninety-eight percent.
My first reaction was disbelief. My second was the one most working developers will have: I have spent enough late nights inside old codebases to know how big that promise is, and how many ways it can go wrong. So I dug into what was actually reported, what is still unproven, and what the rest of us should take from it.
What actually happened
Fortune first reported this on October 30, 2026: Cisco president and chief product officer Jeetu Patel has given Padmanabhan Krishnamurthy, a 27-year-old engineer who goes by Paddy, unlimited AI tokens to refactor the company's legacy code. Patel's own words were: "I've given him unlimited tokens because he's going to refactor code." Per Tom's Hardware, he added that Krishnamurthy could "go nuts."
Krishnamurthy is not a random junior hire. He co-founded EzDubs, a Y Combinator-backed real-time translation startup Cisco acquired in 2025 and folded into Webex. Inside Cisco, he ended up five levels below Patel, in a role Patel called "unstrategic," until Patel heard his pitch during a visit to the San Francisco office and backed it hard.
The pitch is the interesting part: Krishnamurthy told Patel a team of two could refactor Cisco's legacy code with AI agents, targeting 98% of it within 24 months. But first, he gave himself four weeks to prove the approach works at all. If those four weeks show the method holds up, the longer effort gets its runway. If they do not, it does not.
That four-week proof window is the most important detail, and the one most headlines skipped. Nobody has reported any results yet: no completed migrations, no test coverage numbers, no defect rates. The 98% figure is a target somebody set, not a result anybody achieved.
The fine print: 98% is a target, not a result
Here is what we do not know, and the gaps are enormous: which code he is rewriting first, which AI models or tools he is using, how the four-week proof period is being judged, or even when it started. And nobody has said what the 98% is 98% of: lines of code, services, repositories? Without a denominator, the percentage is a slogan, not a measurement.
This is basic discipline, not cynicism. When I refactor a WordPress site, the first thing I do is define what "done" looks like: which pages, which templates, which behaviours must stay identical. If I cannot name the scope, I cannot honestly claim a percentage of it is finished.
Coverage is also split between "rewrite" and "refactor," very different jobs. A rewrite replaces code with new code that hopefully does the same thing. A refactor restructures existing code while keeping its behaviour the same. The safer reading, and the one Patel's quote supports, is refactoring: modernising code in place with AI agents, piece by piece, without changing what it does. Still a huge job, but far more sane than rewriting a networking giant's codebase from scratch.
What unlimited tokens actually cost
The phrase "unlimited tokens" sounds like a blank cheque, and in a sense it is. But the tokens are the cheap part of this experiment. Checking the code is the expensive part, and that is where the real cost hides.
The only hard cost data point I found: an InfoQ case study from September 2026, where AI agents refactored a C codebase of around 300,000 lines in three weeks for about $4,000 in tokens. A real number, strikingly cheap for the amount of code involved.
Now scale that up. Cisco's own blog said in 2021 that IOS XE, the software running many of its switches and routers, contained around 190 million lines of code. Scaling that to 190 million lines lands around $2.5 million in tokens: my rough maths on someone else's published numbers, not a fact about Cisco's bill. A few million dollars to touch a codebase that size is a rounding error for Cisco.
The real bill is people. Somebody must review what the agents produce, maintain the tests proving the new code behaves like the old code, chase down subtle behaviour changes, and take responsibility when something breaks in production. The team stays tiny. The machines do the typing. The humans do the judging.
Why Cisco is doing this now
This did not come out of nowhere. At Cisco Live in June 2025, Patel said he wanted Cisco's engineers to write half as much code by hand and hand routine work to AI agents. This experiment is that strategy with the training wheels off.
There is hard business logic behind it too. Legacy code is not just ugly code, it is slow code. Every new feature must be built around quirks nobody fully remembers. Every security patch takes longer because the engineer must first understand ten-year-old decisions. Every new hire takes months to become productive. I have felt this with old WordPress themes and plugins, where one badly written function from five years ago can eat an entire afternoon. Multiply that by a codebase the size of a small country's library, and leaving the code alone starts to cost more than rewriting it.
Cisco can afford the experiment. In May 2026 it reported record quarterly revenue of $15.8 billion, up 12% year over year, while announcing it would cut fewer than 4,000 jobs, under 5% of its workforce, in the following quarter. The picture: Cisco is spending on AI while trimming headcount elsewhere, and this experiment is part of that same bet.
What this means for developers like us
I keep coming back to one question: if two people with unlimited tokens can plausibly take on a codebase of Cisco's size, what does that mean for the rest of us?
First, the bottleneck in software is moving. For most of my career, the bottleneck was writing code. AI agents have already made that dramatically cheaper. The bottleneck now is knowing what the code should do, and proving that it does it. Tests, specifications, and careful review matter more than ever, and the developers who thrive will be the ones who can define correct behaviour precisely, not the ones who type fastest.
Second, small teams are about to punch far above their weight. Krishnamurthy's pitch was not "give me five hundred engineers." It was "give me two people and no cap on tokens." If that works at Cisco's scale, it works at every scale below it. A freelancer with a good AI workflow can now take on jobs that used to need an agency. That is exciting, and it is also a warning: developers who refuse to learn these tools will compete against developers who use them fluently.
Third, the boring parts are where the value is. The glamorous version is an AI writing thousands of lines of fresh code. The real version is unglamorous: writing tests for old behaviour, documenting edge cases, reviewing diffs line by line. I have learned this the hard way on my own projects. The AI can generate a beautiful new version of a function in seconds. It cannot tell you the old version had a quirk three other plugins depend on. Only a human who understands the system can catch that, and only tests can prove it.
The risks nobody is talking about
I am optimistic, but the risks deserve spelling out. The biggest is silent behaviour change. Legacy code is full of undocumented behaviour other systems depend on. An AI agent refactoring a function can easily "fix" a quirk that turns out to be load-bearing. The code compiles, the tests pass, and six months later a customer discovers a feature they relied on quietly changed. The only defence is thorough testing before and after, and that testing is slow, human work.
The second risk is security. Old code often contains old vulnerabilities, and refactoring can introduce new ones. AI-generated code needs the same security review as human-written code, ideally more: it can produce plausible-looking code that handles edge cases badly. Cisco's code runs networks the world depends on. The stakes are not theoretical.
The third risk is the one I worry about most for our industry: the temptation to believe the headline number. If Cisco's experiment is seen as a success, every executive in tech will want their own 98% target. Some will be done carefully, with tests and review. Others will be cost-cutting exercises with review skipped. The difference between those two versions is the difference between a modernisation success story and a production outage.
My honest conclusion
So what do I actually think will happen? I think the four-week proof period will show AI agents can refactor real legacy code much faster than a human team. I also think the 24-month, 98% target will slip, get redefined, or quietly turn into a smaller number with a clearer denominator. That is not failure. That is what honest engineering looks like when it meets an ambitious target.
The real story is not the number. It is that one of the biggest technology companies in the world has decided the cheapest way to deal with its legacy code is to give a small team unlimited AI and get out of their way. Developers who learn to direct AI agents, write the tests that keep them honest, and review their output carefully will be the ones in demand. The typing was never the job. The judgement was.
I will be watching what comes out of those four weeks, and if Cisco publishes results, I will write a follow-up. For now, the headline is exciting, the fine print is humbling, and the lesson is simple: the future belongs to developers who can tell an AI exactly what "correct" means, and then prove it.
Frequently Asked Questions
Who is the engineer Cisco gave unlimited AI tokens to?
Padmanabhan "Paddy" Krishnamurthy, a 27-year-old engineer. He co-founded EzDubs, a Y Combinator-backed real-time translation startup Cisco acquired in 2025. He worked five levels below Cisco president Jeetu Patel, who heard his pitch and backed it.
What is the actual goal of the Cisco legacy code project?
The stated target: refactor 98% of Cisco's legacy code within 24 months using AI agents, with a team of two, after a four-week window to prove the approach works. No results have been reported yet, so 98% is a target, not an achievement.
What does "unlimited AI tokens" mean?
AI models charge by the token, roughly a chunk of text sent to or generated by the model. "Unlimited tokens" means Cisco removed the spending cap, so Krishnamurthy can use as much AI compute as the work needs. The tokens are relatively cheap; the expensive part is human review and testing of what the AI produces.
How much would it cost to rewrite that much code with AI?
Unknown, because no model or tool has been named. The only published comparison I found is an InfoQ case study from September 2026: AI agents refactored a 300,000-line C codebase in three weeks for about $4,000 in tokens. Scaling that to Cisco's 190-million-line IOS XE codebase gives roughly $2.5 million, but that is an estimate, not a reported fact.
Can small teams or freelancers do the same thing with AI coding agents?
Yes, in principle. The tools behind this experiment are available to everyone. A freelancer or small agency can already use AI agents to refactor WordPress themes, plugins, and old custom code far faster than by hand. The catch is the same one Cisco faces: you still need tests, careful review, and a human who understands what the old code was supposed to do.
