Daily Digest — 2026-08-24

Andrew Ng's team at deeplearning.ai conducted a comprehensive study to determine the essential skills for AI engineers in the evolving landscape of software development. Their analysis involved over 10,000 job postings and interviews with AI experts, hiring managers, and recruiters. This research culminated in the identification of four major AI engineering skills, which are presented as critical for staying relevant as software development integrates more with AI.

1. Building and Deploying AI Applications: This skill goes beyond merely calling an LLM API. It encompasses a deep understanding of concepts such as context engineering, which involves crafting effective prompts and managing conversation history for LLMs. It also includes knowledge of Retrieval Augmented Generation (RAG), where external knowledge bases are used to improve AI responses, and working with AI agents that can autonomously perform tasks. Furthermore, an understanding of machine learning fundamentals is important. Critically, engineers need proficiency in evals and error analysis, as AI outputs are inherently unpredictable, unlike traditional software. This necessitates rigorous testing and debugging strategies tailored for probabilistic systems.

2. Software Engineering Fundamentals: This skill is emphasized as a counter-narrative to the idea that coding agents diminish the value of traditional engineering knowledge. Ng argues that a strong foundation in software architecture, reliability, security, data management, scalability, and trade-offs is more important than ever. These fundamentals enable engineers to effectively guide and steer AI coding agents, ensuring that the generated code is robust, efficient, and maintainable. Without this understanding, engineers may not even recognize when an AI agent is making suboptimal or incorrect decisions, leading to potentially critical issues in production.

3. Using Coding Agents Effectively: This skill extends beyond basic prompting. It requires a nuanced understanding of context management within agent interactions, knowing when to instruct the agent to plan versus execute a task, and how to implement verifiers that allow agents to autonomously validate their own outputs and close feedback loops. Engineers also need to understand when and how to utilize specifications to guide agent behavior and, for complex projects, how to coordinate multiple agents to achieve a larger goal. This involves orchestrating different AI components to work together seamlessly.

4. Product Sense: This is highlighted as potentially the most significant shift. As AI agents become more capable of implementing technical specifications, the engineer's value moves upstream to the decision-making process. This means engineers need a strong product sense, including an in-depth understanding of the customer and business context. They must be adept at identifying opportunities where AI can provide significant value and making critical decisions about when to rapidly prototype ideas versus when to slow down and carefully engineer solutions for production. This shift demands that engineers contribute more to strategic product definition and less to rote coding tasks.

Uber's presentation at the AI Engineer World's Fair outlines their strategy for a managed software factory powered by Agentic AI. Their goal is to streamline the entire Software Development Life Cycle (SDLC) using AI agents, from initial ideation to production and maintenance. The results have been impressive: a 78%+ leverage of PRs by agents, a 2x increase in lines of code per engineer year-over-year, and over 250 automated migrations collectively handling 9 million lines of code.

The foundation of this software factory is built upon six interconnected blocks, forming their Cortana Platform:

1. Model Gateway: This serves as the single, secure entry point for all AI model requests, both internal and external (from vendors like OpenAI, Anthropic, Google, AWS). Key considerations for the Model Gateway include ensuring no PII ever leaves Uber's perimeter, enforcing strictly bounded latency (under 100ms), and enabling cost attribution per user, project, and team. It features middleware for identity (using Spire), data anonymization (redacting 20+ PII types), an AI Guard with five specialized safety and policy models, and caching/token optimization strategies. Currently, it handles over 100 million model requests daily across 800+ projects.

2. MCP Gateway: This gateway provides a unified interface for agents to access thousands of internal and third-party SaaS APIs (e.g., Slack, Google, Jira). Previously, diverse authentication and setup processes for each tool caused significant token tax and operational overhead. The MCP Gateway offers a consistent experience with features like authentication, discovery, schema handling, routing, and execution. An automated crawler identifies internal APIs and projects them into MCPs (Managed Control Plane) with a single configuration change. For SaaS tools, Uber hosts them and handles token exchange, simplifying integration. They've also developed Omni MCP, a single MCP that can discover and invoke any other MCP. By projecting MCPs into CLI patterns, the response context footprint is reduced, saving over 40% fleet-wide. A code-mode skill dynamically creates Python scripts for hyper-optimizing top MCP token-consuming use cases.

3. DevPods: This is Uber's cloud-based remote development environment, extended to support AI agents. DevPods offer pre-provisioned Kubernetes Balloon Pods, allowing agents to start working in seconds with pre-built repositories, snapshot stores, and search indexes. This significantly speeds up agent execution by eliminating environment setup time. They've also introduced a Mega DevPod that consolidates all repositories into a single environment, crucial for autonomous coding agents that need to work across multiple codebases. Additionally, they provide one-click setup for non-engineer access to agent environments, blurring the lines of traditional engineering roles.

4. Agent Skills: This block addresses the challenge of managing diverse, often duplicated, and inconsistent skills developed across teams. Uber has built an entire lifecycle around skills, offering core skills and domain-specific skills through a managed skills marketplace (currently 2,500+ skills). Skills undergo lint checks and automated reviews to ensure a baseline quality. A single command facilitates plugin discovery and installation. Based on engineer personas, default skills are auto-installed for agents. A feedback loop is established through traces and comments, enabling continuous skill improvement. Over 20,000 skill executions occur daily.

5. Context Graph: To address agents spending excessive time and tokens on basic context retrieval, Uber developed a Context Graph. This graph integrates 150+ unique node and edge types with 40 million+ entries, representing Uber's entire operational knowledge. It connects mobile app builds, backend services, data lakes, design docs, Jira tickets, and incident bugs. This allows agents to quickly find relevant context within the ecosystem, significantly improving token usage, turns, and latency. Early evaluations show massive improvements with the graph's integration across planning, ownership, data analysis, and security use cases.

6. AI Assistant (Cortana): Cortana is Uber's unified AI assistant, integrating all the aforementioned building blocks. It provides a common surface for all employees, accessible via Slack, CLI, or web. It allows for personalized experiences, where users can hook up custom skills and prompts, and integrate with their team's Slack channels. This enables collaborative AI workflows. In the last month alone, 300+ unique personas were created, resulting in over 20,000 daily sessions.

Uber's agentic AI transformation is still in its early stages, but these building blocks provide a solid foundation for further acceleration. Future investments will focus on addressing bottlenecks in infrastructure scalability, experimentation, and decision-making to continue advancing their managed software factory.