Amazon warns against tokenmaxxing, telling engineers that using more AI doesn't necessarily mean they're doing better work

Amazon is telling its software engineers to be more selective about AI. In a memo sent Tuesday, the company said it has entered “the era of the Renaissance Developer,” where engineers mix AI with their own judgment to build new things rather than lean on tools blindly.

The sharpest line in the memo is also the simplest: if AI doesn’t help a task, that isn’t a failure. Terk Terkowitz, vice president and distinguished engineer at WorldWide Amazon Stores, told engineers to document that, share it with peers, and bring it to managers. The point is not to shame people for skipping AI. It’s to treat that judgment as part of the job.

Amazon is pushing back on a very modern kind of nonsense: measuring productivity by tokens and lines of code. Terkowitz called those metrics deeply flawed, since they can miss whether the work is actually good. Turn them into targets, he warned, and you get higher costs, more complexity, and less of Amazon’s “invent and simplify” mindset. He even gave the practice a name: “tokenmaxxing,” which he said clashes with frugality.

That concern isn’t theoretical. Amazon has already had to shut down an internal AI leaderboard called Kirorank after employees reportedly used AI agents on pointless tasks just to inflate their token totals. The company pulled the plug in May because the stunt raised concerns about unnecessary computing costs, which can pile up fast.

Amazon is still very much in the AI camp, though, and it hasn’t thrown out its adoption metrics. Terkowitz said they still “correlate well with the outputs that matter.” Managers are being asked to compare low-usage and high-usage employees: why some use AI little, and whether the heavy users are actually producing more. The message is not anti-AI. It’s anti-fetish.