## Sources

1. [Stop Getting Good at Protocols. Get Good at Agent Experience.](https://www.oreilly.com/radar/stop-getting-good-at-protocols-get-good-at-agent-experience/)

---

### **Stop Getting Good at Protocols. Get Good at Agent Experience.** 
**Author:** Sean Roberts

This article argues that the technology industry is currently caught in a cycle of "hype and rewrite" regarding AI agent protocols, much like it previously did with microservices and GraphQL [1, 2]. The author advocates for shifting focus from mastering specific technical protocols to developing a disciplined practice of **Agent Experience (AX)** [3].

#### **Main Arguments**
*   **The Tool Trap:** The industry has a habit of confusing specific tools or protocols (like MCP, AI Skills, or A2A) with actual strategy [2]. Building a strategy around a protocol means building on a foundation controlled by others that the market may abandon at any moment [3].
*   **AX as a Design Problem:** Similar to how User Experience (UX) emerged as a design discipline that transcended specific UI frameworks like React, Agent Experience is a discipline focused on how agents discover, understand, and interact with systems [4].
*   **Protocols are Secondary:** A protocol is merely a "serialization format" or "implementation detail" [5, 6]. The real challenge lies in the **experience design**—understanding what an agent needs to accomplish and providing the necessary context for it to succeed [5].
*   **Preventing Silent Churn:** If an agent struggles with an API—due to poor authentication, high token usage, or lack of context—it will quietly switch to a competitor's service that offers a better experience. The human customer may never even realize why the switch happened, leading to lost business without a single support ticket [7].

#### **Key Takeaways**
*   **AX is an Extension of Existing Disciplines:** AX does not replace UX, Developer Experience (DX), or Customer Experience (CX); rather, it extends them to account for the agents that customers now delegate tasks to [7, 8].
*   **The Protocol Treadmill:** Constantly chasing the latest protocol (from MCP in 2025 to AI Skills in 2026) is a "treadmill" that only speeds up. Teams should instead focus on the underlying problem space which remains constant regardless of the toolchain [5, 9].
*   **Evidence of Poor Design:** A 2026 study of 103 MCP servers found that **97.1% of tool descriptions had quality issues**, and 56% failed to state their purpose clearly. This suggests that while the protocols worked, the experience design was fundamentally flawed [10].
*   **Strategic Advantage:** Organizations that build an AX practice now will outperform competitors who are merely performing successive protocol migrations [11].

#### **Important Details & The AX Practice**
The author outlines a five-step practice for taking Agent Experience seriously [12]:
1.  **Audit the Agents:** Use traffic data and logs to identify which agents (e.g., Claude Code, Cursor, or custom APIs) are visiting your system [12].
2.  **Identify Use Cases:** Determine which high-value tasks customers are attempting to delegate to these agents [12].
3.  **Verify and Audit Interactions:** Perform "usability testing" for LLMs to see where agents get stuck, misunderstand services, or lack context [11].
4.  **Improve and Repeat:** Treat AX as a living practice. For example, if agents consistently assume a product works a certain way, it may be more effective to change the product to match agent expectations than to fight those assumptions [11].
5.  **Automate Validation:** Use frameworks like **AXIS** (an open-source scoring framework) to run agents against real scenarios and catch AX regressions in the CI/CD pipeline [13].

The article concludes by noting that while learning a new protocol is a bounded technical problem with a clear finish line, building an AX discipline involves sitting with **ambiguity** and studying behavior [14]. However, with 35% of organizations already using agentic AI as of 2026, the shift from UX to AX is becoming a primary engineering concern [15].