Serverless Panel • N. Coult, R. Kohler, D. Anderson, J. Agarwal, A. Laxmi & J. Dongre
A panel of experts from AWS, G-P, and AntStack discuss the practical impact of Generative AI on software development. They explore how AI is used as an accelerant for productivity, the challenges of applying it to large-scale system design, its role in modernization, and the future implications for developer careers and safety-critical systems.
AI as a Productivity Accelerant, Not a Replacement
The panel unanimously views AI not as a replacement for human intellect but as a powerful accelerant. Nick Coult from AWS uses AI daily to explore capabilities and boost personal and team productivity. He notes that AI transforms workflows, enabling tasks that were previously too time-consuming.
“I can have a prototype up and running in 15 minutes. Whereas before AI, I wouldn’t have done that… That’s not just acceleration. That’s actually doing something that I wouldn’t have done before because I just didn’t have the time.” - Nick Coult
This sentiment is echoed across different roles:
- Product Management: Janak Agarwal (AWS) described AI as a “forcing function” that helps product managers keep pace with the 10x increase in requests from AI-enabled customers and sales teams.
- Technical Leadership: Robbie Kohler (Byte by Yum!) uses GenAI to stay connected to the codebase by having it explain PRs, allowing him to maintain technical oversight without getting lost in the weeds. It also helps him quickly understand code in unfamiliar languages like Go or Rust.
- Development: Akshatha Laxmi (AntStack) explained that AI has “significantly shortened the scaffolding time” for new applications. It helps maintain organizational coding standards and is particularly useful for identifying edge cases during testing that a developer might overlook.
Challenges of AI in Large-Scale System Design
While AI excels at implementation, its role in architecting massive, mission-critical systems is still limited. Nick Coult highlighted the sheer scale of AWS Lambda, which handles over 10 trillion function invocations per month.
- Training Data Limitations: LLMs are trained on public data like GitHub repositories, which do not contain examples of systems designed for Amazon’s scale. Therefore, an LLM cannot be trusted to design a system like AWS Lambda.
- Human-Driven Architecture: At AWS, humans drive the high-level architecture and the most complex 10% of the code. AI’s primary role is to fill in the boilerplate and standard patterns, accelerating the implementation phase. A senior principal engineer could prototype a complex system in Rust in a couple of hours, a task that would have previously taken weeks.
Modernization, Refactoring, and Customer Feedback
The panel discussed how AI changes the approach to modernization and how a company like AWS handles its own internal evolution.
- AI-Powered Modernization: David Anderson (G-P) reflected that if GenAI had been available during past modernization efforts, it would have significantly helped by scaling best practices, accelerating learning through a “perpetual hackathon” model, and improving the quality and robustness of the systems being built.
- AWS’s Approach to Change: AWS prioritizes backward compatibility, making large-scale refactoring of live services a rare and lengthy process (e.g., a five-year rewrite for a critical ECR backend). It’s more common for AWS to launch a new, improved service rather than refactor an existing one in a way that could impact customers.
- Handling Customer Feedback: Janak Agarwal explained that product strategy is not just about listening to feature requests but triangulating them with the business strategy. Feedback is clustered into themes, and the team prioritizes solving a complete “arc” of related problems to fully unblock a customer’s workload, rather than addressing isolated requests.
The Future of Development with AI
The discussion addressed the long-term implications of AI on the software development profession.
- Skill Atrophy vs. Skill Enhancement: An audience member raised the concern that over-reliance on AI could lead to the atrophy of fundamental coding skills. While acknowledging the risk, the panel emphasized that AI primarily automates boilerplate, and developers still need to read, write, and understand code. The bigger question is how the career path from junior to senior engineer will evolve when entry-level tasks are increasingly automated.
- The Need for Human Oversight: Akshatha Laxmi stressed a key philosophy: “AI is as good as the person who’s using the AI.” A human with strong fundamental knowledge must always review AI-generated code to catch logical errors, performance issues (e.g., a Big O of n-cubed algorithm), and business logic bugs.
- Observability and Safety: In the future of agentic AI, observability will be crucial. Nick Coult predicts that 90% of observability principles will remain the same (latency, error rates, cost), but new metrics specific to AI, like token usage and handling non-determinism, will become important. For safety-critical systems (e.g., aviation, military), AWS is developing capabilities like guardrails for Bedrock that use automated reasoning—a formal computational logic—to verify the correctness of LLM outputs, ensuring reliability in high-stakes environments.