Amazon warns engineers against tokenmaxxing to prioritize quality over volume

The rise of tokenmaxxing

Amazon recently released an internal memo to software engineers that addresses a specific trend in the corporate AI rollout. The company is actively discouraging a behavior it calls 'tokenmaxxing,' which occurs when developers use artificial intelligence tools as much as possible without considering the quality of the results. This behavior often stems from a rush to adopt new technology quickly.

As companies integrate AI into their workflows, they have faced challenges regarding the surging costs of these models. Tokens are the basic units of data that AI models process and generate. When employees use these tools excessively for simple tasks, it can lead to inflated bills and unnecessary complexity in the software being built. Some companies have already responded by setting hard limits on token usage or switching to cheaper open-source models.

Amazon has specifically noted that metrics like token usage and the rate of AI tool adoption are 'deeply flawed' ways to measure success. The company warns that these numbers can be misleading if they do not correlate with actual progress. Instead, the memo suggests that engineers should use AI with care and judgment rather than viewing it as a default for every single coding task.

Shifting the productivity metric

To replace raw usage numbers, Amazon is looking toward more meaningful indicators of success. One primary metric the company is highlighting is 'feature velocity.' This measures the actual speed and frequency of successful launches and production changes. The goal is to see how AI helps engineers move projects from conception to completion more efficiently.

Marc Brooker, a vice president and distinguished engineer at Amazon Web Services, explained to CNBC that the company views these metrics as a return on investment. He noted that the company is willing to invest time, human effort, and tokens into problems that matter to customers or the business. The focus is on ensuring that the resources spent on AI actually solve important problems rather than being wasted on trivialities.

By focusing on outcomes, Amazon hopes to avoid the 'wasted effort' that comes from over-automating. The company wants to ensure that the integration of AI leads to a tangible boost in productivity. This shift suggests that the company is moving away from a 'more is better' mindset toward a strategy of intentional, high-impact AI application.

Judging when to skip AI

A significant part of the new guidelines involves telling staffers that it is okay to ignore AI tools when they are not helpful. The memo explicitly states that if a specific task does not benefit from current AI tooling, that is considered a finding rather than a failing. This is intended to encourage engineers to think critically about the limitations of the technology they are using.

Amazon is looking for engineers who can demonstrate the judgment to identify where AI does not yet help their work. The company wants to avoid situations where developers might force an AI solution onto a problem just because the tool is available. This approach requires a deeper understanding of the software engineering process and the specific nuances of the tasks at hand.

This stance reflects a broader corporate realization that AI is still an emerging field. While the company believes AI will be at the center of software engineering practice, it acknowledges that there are still many unknowns. By rewarding engineers for knowing when to step away from the tools, Amazon is prioritizing human expertise and sound decision-making over automated output.

Lessons from internal testing

The move to curb tokenmaxxing comes after Amazon discovered issues with how employees were interacting with internal systems. In May, the company shut down an internal leaderboard after finding that some employees were intentionally tokenmaxxing to climb the rankings. This incident highlighted how easily employees can game metrics that prioritize volume over substance.

The current memo serves as a correction to those types of behaviors. By moving away from leaderboards and usage-based rewards, Amazon aims to create a more stable environment for AI integration. The company wants to ensure that the push for AI adoption doesn't result in a cluttered codebase or a bloated budget for the sake of hitting a number.

Looking forward, Amazon's strategy suggests that the next phase of AI integration will be about refinement. The company has already rapidly integrated AI across its sprawling operations under CEO Andy Jassy. Now, the focus is shifting toward ensuring those tools are used effectively to change the scope and speed of work without creating new problems of complexity or cost.