Six AI experts break down what works in real use: agents, assistants, and the gap between hype

A panel of six AI practitioners—including engineers and product experts—discuss which AI capabilities actually deliver value in daily work versus which fall apart under real-world use. The conversation spans OpenAI's recent DevDay announcements, the distinction between AI models that impress in benchmarks and products that solve actual problems, and the trade-offs between agents (systems that take action autonomously) and assistants (tools that help humans decide).

Key themes include building personal assistants, decision-making models, managing context in long agent loops, coding benchmarks across different models, prompt injection risks when tools have real permissions, and the emerging problem of assistant overload as more AI tools demand a place in workflows. The panel also covers model speed, writing quality, local versus cloud models, and how computer-use capabilities might fit into everyday work.