Skills Are the New SDKs: Rethinking AI Agents

Elvin Aghammadzada, a Data Science Engineer at DataRobot, recently presented a compelling argument that "Skills Are the New SDKs", highlighting critical challenges and emerging solutions in the development of AI agents. Aghammadzada's presentation, delivered at the AI Engineer World's Fair, focused on the intricacies of context engineering, the evolving role of documentation, and the strategic importance of 'skills' in building more capable and reliable enterprise AI.

Understanding the Three Layers of AI Applications

Aghammadzada began by dissecting the architecture of any AI agent into three fundamental layers: the user interface (UI) that the user sees, the system prompt and tool descriptions that the model sees, and the data schema or input/output of tool calls that the data sees. He pointed out that bugs in AI agents often reside in the latter two, less visible layers.

The Context Window Paradox: More Isn't Always Better

A significant portion of the presentation addressed the common misconception surrounding the 'infinite context windows' promised by frontier models. Aghammadzada argued that this perception incorrectly shapes how developers think about Retrieval Augmented Generation (RAG) and Model-Centric Prompting (MCP). The reality, he explained, is that simply dumping vast amounts of data into the context window doesn't guarantee better performance. In fact, he cited a paper on 'context rot' which demonstrates that performance degrades significantly after only 25% of the context window is utilized. This means that as conversations grow longer and more tools are added, the model's performance can decline, mirroring the frustrating experience of seeing code assistants repeat the same mistakes.

The Challenge of Documentation for AI Agents

Aghammadzada also highlighted a growing trend: traffic to documentation websites from coding agents has surged from 10% to 50% year-over-year. This shift presents a unique challenge because traditional documentation is written for human readers who can intuit context, ask follow-up questions, or use search engines to clarify points. AI models, however, lack these capabilities unless a robust 'context engine' is built around them. He noted that even a reasonable parameter name, easily understood by a developer, could lead to hallucination and confidence in incorrect interpretations by an LLM.

Friction vs. Fluency: A New Moat for Enterprise AI

Shifting to the future of enterprise AI, Aghammadzada proposed that 'skills' are poised to become a significant contributor to the evolution of enterprise agents. He drew a parallel to the history of software, where the hardest modes to replicate were the 'friction moats' that made it difficult for users to leave a platform due to high switching costs. However, in the current landscape, he argued, companies are becoming adept at reducing switching costs through better tools and APIs. This is leading to a new type of moat: 'fluency'.

Skills, he explained, create this fluency moat by making the user experience better, compounding value with each added skill. Fluency, in essence, is defined as the satisfaction or minimum friction between intent and outcome. While friction moats are defensive, aiming to prevent users from switching, fluency moats are offensive, making the platform so enjoyable that users genuinely want to stay. He emphasized that the 'feel' and reliability of the outcome, from intent to result, are what skills encode for agents, commoditizing the experience layer.

Teachability: The New Enterprise AI Metric

Aghammadzada introduced 'teachability' as a new, critical item on the enterprise evaluation checklist for AI platforms. Beyond security, compliance, governance, SLAs, and integrations, enterprises now need to assess how easily an agent can learn and apply operational knowledge encoded in skills. The ability for an agent to govern, version, and audit its skills is paramount for first-time or repeat users to quickly adopt and leverage the platform's capabilities.

Standardizing Context: The Path Forward

Drawing parallels to the evolution of web development with HTML and React, Aghammadzada suggested that the LLM and agent ecosystem is currently in a similar nascent stage. He stated that the industry is still piecing together tools and discovering patterns through trial and error, lacking the foundational, systematic patterns that define mature software development. He defined 'context engineering' as the art of dynamically orchestrating information within the context window at the precise moment it's needed, filtering and curating it for optimal LLM performance.

He illustrated this with a diagram showing the 'smart zone' (the first 40% of the context window) where LLMs perform well, and the 'dumb zone' beyond that, where performance degrades. The challenge lies in ensuring that essential instructions, RAG data, execution history, memory, and output schemas fit within the smart zone, preventing the agent from starting in the 'dumb zone' even before a conversation begins.

Skills vs. MCP: Complementary Roles

Aghammadzada clarified the distinct but complementary roles of Skills and MCP (Model-Centric Prompting). MCP defines what actions an agent should take without altering the agent's core definition, while Skills help agents think about problems and customize their behavior. Skills also introduce the concept of self-modification, allowing agents to update their own capabilities based on experience. He noted that MCP serves as a robust, separate server for tasks requiring authentication, access, or significant computational resources, whereas Skills run with the agent's environment, offering a more flexible and efficient approach for reasoning-intensive tasks.

The Ecosystem of Skills

The presentation concluded by highlighting the growing ecosystem around agent skills, with over 26 platforms supporting them, including major players like Claude Code, OpenAI Codex, and GitHub Copilot. He mentioned marketplaces like skills.sh for acquiring and sharing skills, and the emergence of projects like OpenClaw, which showcases the power and risks of agent skills, including self-healing capabilities. He stressed that while LLMs can generate skills, human-authored skills often carry more institutional knowledge and perform better, as LLM-generated skills can sometimes increase token usage and degrade performance.