Cisco CPO: A New Model for AI-Era Engineering Leaders
Cisco's President & CPO, Jeetu Patel, outlined a new mental model for engineering leaders, emphasizing skills like judgment, instinct, and an obsession with safety and unit economics. Patel noted that Cisco is targeting 70% of code to be AI-written by 2027 with superior quality, highlighting the shift from managing human engineers to managing AI agents.
- Cisco has already shipped its first product, AI Defense, with 100% of its code written by AI and anticipates releasing at least six more fully AI-written products by the end of 2026. - The company is shifting from traditional agile development to a "spec-driven development" model, where a team of three humans and five AI agents is expected to triple the output of a traditional eight-person team. - To support this shift, Cisco is developing tools like the Cisco AI Assistant for its Webex Developer Portal, which can generate code snippets, provide instant access to documentation, and guide developers through its APIs and SDKs. - Jeetu Patel has emphasized that the primary constraints to this AI-driven future are not talent, but potential shortages in power and data center infrastructure, a growing "context gap" in AI systems, and a "trust deficit" that requires robust security. - The focus on unit economics stems from the different cost structure of AI applications, where marginal costs related to inference, API calls, and compute are significant, unlike traditional SaaS models. - Patel anticipates that AI will make individual engineers 10 to 50 times more productive, reducing the time to turn an idea into a product from months to minutes. - To govern the use of AI agents, Patel suggests they need "background checks" similar to employees, a concept Cisco is backing with billions in security infrastructure to protect agents from attacks and prevent them from "going rogue." - This strategy is supported by hardware innovations like the Cisco Silicon One G300, a switching silicon designed to power large-scale AI clusters and improve job completion times for training and inference workloads by 28%.
