In late 2025, AI models from OpenAI and Anthropic crossed a critical threshold in reliability for coding tasks. This transformed AI coding agents from experimental tools that required constant supervision into practical systems capable of independently building functional applications from high-level prompts.
AI coding agents are automating the manual task of writing code, elevating the engineer's role to that of an architect, prompter, and reviewer. The most effective engineers are now those who can manage multiple AI agents in parallel, rapidly prototype ideas, and critically evaluate AI-generated output.
Companies are beginning to experiment with a model where engineers do not write code directly but instead manage a swarm of AI agents that do the building. This approach, exemplified by StrongDM's experiment, treats code generation as a cheap, automated commodity, focusing human effort on specification and quality control.
The rapid advancement and adoption of AI coding agents create a paradox: while productivity soars, so does the risk. The speaker predicts a "Challenger disaster of AI," where institutional overconfidence and the use of agents in unsafe modes will inevitably lead to a catastrophic failure.
Keep pulling the thread on Simon Willison.