Cisco gives engineer unlimited AI tokens to rewrite legacy code

Original Article ↗ Hacker News Discussion ↗

Cisco’s plan to let a 27-year-old engineer use “unlimited” AI resources to rewrite 98% of its legacy code is being treated as both a bold bet and a potential catastrophe. Commenters question how AI can safely replicate decades of accumulated bug fixes, undocumented edge cases, and interoperability quirks in mission-critical networking software, and whether one relatively junior person can hold the necessary domain knowledge. Others argue modern LLMs make large-scale refactors newly feasible—especially for internal tools or well-specified components—but warn that testing, regression coverage, and clear goals matter far more than flashy claims about unlimited tokens.

Scope and Goals of the Rewrite

  • Many commenters note the article never clearly states why “98% of legacy code” must be rewritten or what concrete outcome Cisco wants (performance, security, maintainability, feature parity, consolidation, etc.).
  • It’s unclear what “legacy code” covers: internal tools, product firmware (e.g., IOS/IOS-XR), specific lines (switches, firewalls, collaboration), or all of the above.
  • Several people worry that touching core networking firmware could have serious stability and interoperability consequences.

Feasibility of AI-Driven Rewrites

  • Enthusiastic voices say late‑2026 LLMs can already translate and refactor large codebases, reverse‑engineer old binaries, and generate huge volumes of code and tests very quickly.
  • Some report success rewriting embedded/legacy systems in months with modest token spend plus a mid‑level developer supervising.
  • Others argue this task is vastly harder than the CEO narrative suggests, especially for complex systems like Cisco’s network OSes.

Legacy Behavior, Bugs, and Testing

  • Strong concern that legacy quirks, bug‑as‑feature behaviors, and interoperability hacks (with other vendors, weird hardware, subtle protocol edge cases) are undocumented and hard to rediscover.
  • Multiple people stress that the real challenge is not code generation but:
    • Capturing existing behavior via extensive regression and end‑to‑end tests.
    • Deciding which bugs to preserve vs. fix without breaking customers.
  • AI is seen as good at:
    • Writing tests and documentation.
    • Refactoring and explaining code.
      But weaker at knowing which behaviors matter or designing truly effective test suites without expert guidance.

Reliability, Risk, and Customer Impact

  • Many worry this could destabilize infrastructure customers rely on for high reliability, unlike a typical SaaS.
  • Some suggest parallel operation, heavy simulation, or A/B rollouts; others fear customers become the de facto test bed.
  • There’s skepticism that “unlimited tokens” and a small team can safely replicate decades of accumulated, hard‑won bug fixes.

Organizational and PR Dimensions

  • Several see the story as a PR/hype piece: emphasizing “AI-native” youth, “unlimited tokens,” and hero rhetoric more than engineering detail.
  • Some read it as a signaling move to investors (“we are very AI”) or a prelude to workforce reduction.
  • There are critical side discussions about Cisco’s historic software quality, expanding product line, and customer frustration, leading some to doubt this will improve matters.