Why agentic AI needs better experts (www.spinellis.gr)

In a recent exploration of agentic AI capabilities, a developer updated the sed program within the uutils project to improve compatibility and performance by switching the default handling of data from characters to raw bytes. Collaborating with OpenAI Codex, the developer experienced both the agent's strengths in managing tedious tasks and its weaknesses in producing production-quality code. Despite Codex efficiently compiling the modified code through a well-structured set of prompts, the final output still required careful review and significant expert guidance—61 out of 78 prompts requested improvements. This case highlights the ongoing need for human oversight when utilizing AI agents like Codex, particularly as their applications become more complex. The session illustrated that as coding agents evolve, the expectations for their output will increase, necessitating even more rigorous expert input. Additionally, the environmental and monetary costs associated with AI usage were significant, with reported carbon emissions and water usage comparable to moderate travel. This invites the AI/ML community to critically assess the balance between efficiency gains from AI automation and the broader implications of their use in software development.