Unlimited tokens for a 98% legacy-code rewrite: what would count as success?
Collected summaries describe Cisco giving an engineer an unlimited token budget and a target of rewriting or refactoring 98% of legacy code in 24 months. A budget and a target are not delivery results; scope and business behavior determine what completion means.
What to consider
Start with an acceptance list, then decide whether to expand AI output. Two summaries of the same story say Cisco president and chief product officer Jeetu Patel assigned a 27-year-old engineer to legacy modernization, with an unlimited token budget and a target of refactoring or rewriting 98% in 24 months. They provide no achieved completion rate, production-quality results or maintenance costs.
Define what the 98% target measures
Lines, files, modules and business capabilities yield different completion rates. Define scope, external consumers and excluded work before choosing a denominator. Record a new implementation, code review and a production switch as separate milestones.
The summaries concern one event, not two independent replications. The 98% and 24-month figures describe a target. They establish neither a monthly delivery schedule nor evidence that rewritten business functions work correctly.
Compare old and new behavior on a real business path
Choose a bounded business path whose old implementation can remain available. Record inputs, expected outputs, failure behavior and dependencies. Run both implementations against the same checks and classify differences as intended fixes, requirement changes or accidental regressions.
Before switching, check whether consumers still receive the behavior they expect and whether a rollback works. State-bearing components need separate migration and recovery checks. A successful compile does not justify overwriting existing records. These are suggestions, not claims about Cisco’s architecture.
Find the bottleneck in delivery records
Separate generation time, human review, test queues and regression fixes. If output accelerates while unreviewed changes pile up, reduce batch size and improve business examples. More simultaneous generation is not automatically the next useful step.
A target and resource allocation do not establish labor savings or suitability for every legacy system. We did not read the original reporting, run Cisco code or measure token use. The practical takeaway is to make modernization comparable and reversible, rather than copy a headline target.
Source links
The links below were supplied with the collected records. This article offers independent analysis of their summaries; we have not individually checked the original sources or tested the products.