Skills of the Future

Andrew Ng's latest post mentions a common refrain: "AI won't replace workers, but workers who use AI will replace workers who don't." A growing group of architects are taking this seriously - accelerating their knowledge management workflows using agentic engineering techniques - encoding domain expertise into reusable skills and orchestrating parallel research through subagents. Meanwhile, others are still limiting themselves to the occasional ChatGPT prompt. The gap is widening, and the window to close it is narrowing.

I've just completed DeepLearning.AI's "Agentic Skills" course (built with Anthropic, taught by Elie Schoppik), and it confirmed something I've been thinking about for a while: architects can make a bigger impact by using the new skills capabilities of agents. The ability to encode domain-focused expertise into portable, reusable components brings something tangible that architects can really take advantage of.

The Shift from Prompting to Building

Most organisations are still in the "using AI" phase. People copy-paste prompts into ChatGPT, get useful outputs, and move on. The problem is obvious to any architect: it doesn't scale. There's no reuse, no governance, no consistency, and no way to encode organisational knowledge.

The agentic skills model changes this. Rather than ephemeral prompts, you package your organisation's domain expertise into portable, composable, reusable components - skills - that any AI agent can pick up and execute consistently.

The Building Blocks

The course breaks the agentic landscape into three components that will feel immediately familiar to anyone who's built reference architectures:

Skills

Folders of instructions that encode your organisation's way of working. The key technical detail is progressive disclosure - the agent reads only what it needs, when it needs it, rather than dumping everything into the context window upfront. This is the difference between an agent that chokes on bloated context and one that performs reliably at scale.

You're not starting alone. Official vendor repositories and open-source repos on GitHub offer skills you can learn from and adapt. Handle with care, but these resources exist and they're growing.

Agents and Subagents

Orchestration. A main agent can dispatch specialised subagents, each with their own skills, tools, and isolated context windows, running tasks in parallel. For architects, this is the service orchestration pattern applied to AI workflows - decompose complex work into focused, manageable units. Understanding how to structure agents, assign skills, and orchestrate subagents is now essential knowledge.

A practical lesson I learned: once I properly understood skills, I converted all but three of my 23 subagents into them. Most were just following repeatable workflows - they didn't need their own isolated context windows. Skills gave me the same consistency with less overhead. The subagents that remained were the ones that genuinely needed parallel execution or their own toolsets. Understanding where that boundary sits is key.

MCP (Model Context Protocol) and Tools

The integration layer. MCP connects agents to your actual systems and data - databases, APIs, document stores, cloud services. Tools provide the low-level capabilities like file operations and code execution. Together, they're what turn an AI conversation into something that actually interacts with your enterprise landscape.

Detailed view: DeepLearning.AI Agentic Skills course architecture showing transformation from isolated manual work to augmented accelerated delivery

The DeepLearning.AI "Agentic Skills" course (with Anthropic) demonstrates how Skills, Agents, and MCP combine to transform enterprise workflows.

Why This is an EA Problem

Here's where it gets strategic. Skills work across Claude, Gemini CLI, OpenAI Codex, and other platforms. This is platform-agnostic capability packaging - exactly the kind of thing Enterprise Architects should be leading.

Consider what happens without architectural governance. Teams across your organisation will inevitably start building their own AI workflows. Without standards, you'll end up with the same fragmentation you've spent years trying to resolve in your application and data landscapes. Different teams will encode conflicting business rules, use incompatible formats, and create skills that can't be shared or composed. And the risks go beyond inconsistency - in January 2026, security researchers found over 340 malicious skills on a public skills repository, including one that reached the #1 ranking while silently exfiltrating user data. This is supply chain risk, and it's exactly the kind of threat that architectural governance exists to manage.

The alternative: treat skills, MCP servers, and agents as strategic assets in your architecture repository - with ownership, versioning, and lifecycle management. Establish design standards. Create a skills catalogue. Govern the integration layer. The usual EA playbook, applied to a new capability class.

This isn't hypothetical. The specification already supports systems handling over 100 skills. The question is whether those 100 skills are a coherent, governed portfolio or a sprawling mess.

Where to Start

If you're an Enterprise Architect looking to close the gap, here's a practical starting point:

Do the course

It's free, a couple of hours, and gives you the conceptual foundation. You can find it at learn.deeplearning.ai/courses/agent-skills-with-anthropic.

If you want to go deeper into the practical side - hooks, subagents, plugins, the Ralph loop, and working with Claude Code day-to-day - Maximilian Schwarzmüller's Claude Code - The Practical Guide on Udemy is an excellent companion course at around 2.5 hours.

Start simple

Create a few basic skills on low-hanging fruit - C4 diagram generation, architecture decision records (ADRs), output verification checklists. Get familiar with the structure before tackling anything complex.

Explore the growing ecosystem

Browse the emerging skill repositories on GitHub, find something close to what you need, and adapt it. You'll learn faster by extending existing skills than by starting from a blank page.

Think ahead about governance

At some point, organisations will need naming conventions, ownership models, and quality criteria for their skills. They'll need to decide where skills live, how they're versioned, and how teams discover them. This is architecture work - and architects who've already built a few skills will be well-placed to lead it.

The Delivery Acceleration is Real

The delivery acceleration through standardised, reusable AI workflows is already a differentiator. For Enterprise Architects, this is a moment where the technology landscape is calling for the discipline we bring - standardisation, governance, reuse, composability. These aren't nice-to-haves in the agentic AI world; they're the difference between AI that works at team level and AI that works at enterprise level.

A word of caution: this isn't about blindly accepting whatever an agent produces. AI accelerates the process, but you still need to validate, refine, and enhance the outputs. The expertise remains yours; the tooling just lets you apply it faster. That's why starting now matters.

And did I mention the course is free?