iko-azoulay (EVP Technology at Salt Security) recounts rolling AI out across R&D. It started with GitHub Copilot licenses (and arguments with the CFO), moved through a three-month “accelerator” trying every tool, and settled mainly on Claude Code for its codebase understanding and plan mode. Copilot stays for tab completion as a “safety belt”. Adoption levers: a GenAI guild of early adopters sharing success stories and workshops, and a challenge to spend a sprint writing no code by hand. The mindset shift is the real hurdle. The flow: a detailed spec as a contract between human, machine, and product; tech design grounded in repo context; task breakdown; AI validating AI at every step, while humans stay accountable. Next steps: product creating working-mock PRs on real data, hierarchical cross-repo context, and less rigid, more interactive artifacts. Big picture: “alignment engineering” may be next year’s hype; task lengths agents handle are growing exponentially; Scrum rituals will have to change.
Show: langtalks
Concepts: ai-rd-rollout · ai-sdlc · ai-verification · autonomous-agents-outlook · coding-agent-workflow