Barely a day and a half has passed since I came back from baby bonding leave, and already my thoughts are spinning. Google’s leave policy is generous, and I chose to disconnect completely during that stretch so I could give my family my undivided focus. But, as every developer knows, hitting the reset button often triggers a fresh wave of thinking.
The speed of change in this AI-driven era is frankly breathtaking. One of the first trends I noticed after returning is the emergence of agent skills. In essence, these are packages of instructions, scripts, and resources that agents can locate and use to execute tasks with greater accuracy and efficiency.
The Angular team has also published its own set of Angular agent skills (GitHub), and I have just refreshed my memory on Addy Osmani’s agent-skills repository. His compilation spans the full software development lifecycle and recently attracted a great deal of attention — rightly so. Seeing skills arranged in a coherent structure so that agents can handle complex workflows pulled me back to an idea I was toying with before my break: Agent Building Blocks. To clarify why this interests me, let me revisit one of my favorite projects from last year.
The Experiment: The Angular AI Tutor
Last year, I built the Angular AI Tutor. Its purpose was to teach Angular interactively by walking developers through the construction of a modern application, step by step. You can check it out today through the Angular MCP server.
The surprising twist? Putting it together was not particularly hard in terms of overall complexity. The instructions are open-source and written in straightforward language. It began as a small experiment: Could I get Gemini to teach me Angular?
With a simple 10-line rules file, the answer was definitely yes. Gemini did more than just supply information; it shaped a learning path that emphasized analytical thinking and enjoyment. But as I shifted from “experiment” to “product,” two significant obstacles became clear:
- Lesson Continuity: The challenge of designing a connected lesson sequence where each stage moves toward one navigable application.
- Syntax Recency: The necessity of guiding the agent to adopt modern features (like Signals) that were not part of the model’s original training.
The Maintenance Challenge
Through trial and error, I managed to get past these barriers with a README totaling 824 lines. That was pretty solid for a working tutor! The project drew considerable attention — Angular GDEs were interested in creating their own versions, and teams like Flutter and Firebase were curious about the “Tutor” pattern for their respective ecosystems.
Still, there was a hurdle to tackle first: The logic was far too “coupled.”
The instructions guiding the tutor’s behavior were tightly woven into Angular-specific syntax. That led to several complications:
- Low Reusability: Adjusting the tutor for Firebase or Flutter would have demanded a substantial rewrite.
- Duplication: We ended up with separate rules for the tutor and separate rules for general code generation.
- Update Friction: Introducing a new feature like Signal Forms took extra time because changes had to be applied across several locations.
My initial approach to solving this — splitting the language into separate rules files — did not work out. In the absence of an agent routing mechanism, the model struggled to figure out which “skill” to access and at what moment.
The Vision: Context Pipeline Engineering
In one of my final meetings before bonding leave, I introduced a Router + Building Blocks architecture. The concept was to break the AI tutor’s core into modular pieces that could be reused or swapped for various frameworks. Back then, I was unsure how to bring it to life.
Now that I’m back, I view Agent Skills as a likely enabler of that vision. Yet, some open questions still deserve attention:
- Extensibility: What methods will make these skills adaptable to individual developer workflows?
- Reliability: How do we ensure the skills and their usage by the model stay dependable when other skills, tools, and context are also active?
- Token Optimization: How do we keep the building block pipelines current without expanding the model’s context window too much?
- Stability: As models advance, how do we prevent the “pipeline” from breaking?
I have begun referring to this field as Context Pipeline Engineering. If you are drawn to these deeper, structural aspects of AI engineering, I explored some of these topics in my ng-conf session: Architect or Assembler: Your Role in the AI Future.
It’s good to be back. Be sure to subscribe to our YouTube channel and blog to stay updated as we dig into these topics and beyond. Let’s create something together.
Back from Baby Bonding: Agents, Building Blocks, and the Future of AI Engineering was originally published in Angular Blog on Medium, where people are continuing the conversation by highlighting and responding to this story.