Amazon has updated its operational guidelines for software engineers, cautioning development teams against chasing raw artificial intelligence usage metrics at the expense of actual business results. Terk Terkowitz, Amazon’s vice president and distinguished engineer, outlined the shifting strategy in an internal memo published Tuesday, emphasizing that artificial intelligence is transforming software development more fundamentally and rapidly than prior paradigm shifts like cloud computing.
While Terkowitz stated that engineers must embrace the technological shift and dedicate time to continuous learning to avoid falling behind, he stressed that employee performance should ultimately be evaluated based on tangible impact delivered to customers and the wider business. The updated guidelines mark a significant pivot in how the tech giant assesses its technical workforce and engineering managers as automated coding agents and specialized tools reshape the software engineering profession.
Prior to the revised guidance, Amazon monitored various internal indicators, including raw AI utilization rates, token consumption volumes, and lines of code generated, occasionally employing leaderboards to rank teams against those metrics. However, the memo warns that treating such diagnostic measures as primary performance targets can generate perverse incentives. Specifically, Terkowitz cautioned against “tokenmaxxing,” a phenomenon where teams optimize for high token consumption rather than producing superior software or enhanced customer experiences.
The company acknowledged that tracking metrics such as token usage, tool adoption rates, and automated code reviews can correlate with desired outcomes when interpreted properly, but cautioned that they require careful handling. High token usage can introduce unnecessary operational costs, whereas an overemphasis on raw lines of code can inadvertently inflate software complexity. To navigate these pitfalls, Terkowitz recommended that engineering managers closely examine teams at both extremes of the adoption spectrum: the bottom 10 percent to identify obstacles preventing effective tool utilization, and the top 10 percent to verify whether heavy usage genuinely accelerates product delivery.
According to the memo, the company’s most AI-forward frontier teams have demonstrated measurable improvements in productivity and delivery speeds, and individuals lagging behind in adoption or output metrics risk obsolescence. Nonetheless, the guidance provides allowances for specialized roles or projects where current AI tools offer limited utility. In such scenarios, engineers are expected to document their technical roadblocks, share findings with peers and management, and collaborate toward viable solutions.
Under the updated policy, engineering leaders are expected to directly incorporate AI tools into production workflows, develop a nuanced understanding of their operational strengths and limitations, and articulate that technical judgment to their respective teams. Terkowitz concluded that demonstrating critical thinking about where automated systems fall short is just as valuable as deploying them, noting that an engineer who can clearly explain why AI does not yet benefit their specific workflow is actively displaying the professional judgment the company requires.
