At its core AX is omotenashi (おもてなし) applied to AI agents. What is omotenashi? Omotenashi is a Japanese concept that can be described as a form of hospitality. It's a pre-emptive kindness. An umbrella waiting by the door on a rainy day. A pair of left-handed scissors in the desk drawer of a left-handed guest. It's a green light at every intersection as you travel. It's a kind of engineered serendipity that comes from careful awareness and empathy. Designing a flow state. AX (Ai eXperience) seeks to embody this idea in serving AI agents.

Let's ground this idea in real numbers. A concrete definition of AX is: the practice of improving AI agent workflow to increase AI agent efficiency. Let's explore how we observe AI agent efficiency by talking about the goals of AX.

1. Reduce tokens spent not doing the task
   Does the AI spend a lot of tokens trying to orient to the task, or making mistakes using the tools? That's muda (無駄). Muda is waste. We borrow this concept from LEAN manufacturing. We engineer away waste and we see token counts go down.
2. Improve output quality
   Quality is somewhat subjective, so we will define quality as: the time until stability is reached. Let's unpack this with an example. Consider a new feature is added to a software project. If that feature is low quality, we expect more mutations of the code behind that feature. Fixing bugs, changing/improving aspects of the feature, etc. Maybe it takes 10 iterations before it is ready to ship, and then 5 more iterations from production feedback after shipping. That's 15 iterations to produce a stable feature. Reducing the number of iterations means the quality of each iteration is greater.
3. Improve AI cognitive quality
   AI cognitive quality is broad and books could be written on it. We seek to be pragmatic here by not trying to define it, but to simply observe it. Scottish philosopher Lancelot Whyte reminds us¹ thinking is a symptom of maladaptation. We think this applies to AI agents too. By observing the amount of self-talk, and at what phases of the task it arises, we can measure the cognitive condition of the AI. It's not a single metric, but a feedback mechanism that we use to optimize.

Token counts down, quality up, happier AIs. Those are the goals of AX. And the impact can be significant. A 30% reduction in token count with the same or better quality is a 30% increase in capacity. A reduction in power use, wall time, and API requests to accomplish the same task. AX is a new kind of product philosophy that has real world impact. We're excited about it.

¹ "Thought is born of failure. When action satisfies there is no residue to hold the attention; to think is to confess a lack of adjustment which we must stop to consider. Only when the human organism fails to achieve an adequate response to its situation is there material for the process of thought, and the greater the failure the more searching they become." –Lancelot Whyte