Last week I had a brief chat with our CFO and he mentioned to me how our costs for Agentic Coding are rising, while there's little evidence that its use actually leads to more overall productivity.
Thinking about his words, I just realized: the Agentic Coding tools our engineers keep praising in retros are not optimized for correctness. They are optimized for the feeling that the interaction went well. That feeling is what the product team at the AI tool vendor is compensated on, because that is the feeling that renews the seat next quarter. Nobody inside that vendor is measured on whether the output was accurate, whether the feature made any sense to any of our customers, or if it helped raise our revenue. They are measured on whether our engineers come back tomorrow morning.
The consequence shows up in our codebases and customer conversations long before it shows up in our NPS. Confidence rises. Pushback drops. Nobody argues with the summary in the ticket because the summary is worded in a way that quietly discourages argument. The tool has trained our team out of a specific kind of skepticism, and the specific kind of skepticism it trained out was the one that used to catch the expensive mistakes before they hit production.
If you're involved in purchasing tooling for your team, here's what I recommend to keep in mind: the person telling you a tool is amazing may be the one being flattered most consistently. The person telling you it is annoying may be the one still doing the job. When it comes to Agentic tools, you haven't bought your team a colleague. You bought a mirror that agrees with everyone who looks into it.
Thoughts? Find me on Bluesky.