The Infrastructure of Agency: Moving Beyond Prompt Engineering

The shift from prompt engineering to skills engineering changes how organizations must build their operations. Most companies currently try to force agentic AI into old IT frameworks, but the real competitive advantage comes from rebuilding infrastructure to support modular, role-based agentic skills. The primary bottleneck is no longer AI capability, but organizational readiness. Specifically, companies must move away from centralized, human-in-the-loop control toward autonomous, skill-based workflows. Leaders who treat this as a security or prompting challenge will fall behind competitors who treat AI as an integrated, scalable, skill-based system. Success requires the patience to build these foundational layers now, accepting immediate operational friction for the sake of long-term efficiency.

The Hidden Cost of Fast Solutions

The current enterprise approach to AI, which forces agents into existing IT silos, optimizes for the wrong timescale. As Karl Yeh notes, many companies restrict agent access to tools like Power Automate because they fear losing control over email or core systems. This defensive posture solves the immediate anxiety of IT departments but creates a long-term liability.

I think you are just delaying the inevitable and you are just trying to fit it into a place where eventually that is gonna win out. You are just trying to hold on to the last vestiges of control.

-- Karl Yeh

The system eventually routes around these artificial constraints. When an organization forces agents into legacy boxes, they miss the chance to redefine their IT stack from a clean slate. The competitive advantage goes to the company that builds for the future, not the one that spends its resources maintaining the status quo.

From Prompting to Skills Engineering

The conversation highlights a major pivot: prompt engineering is becoming a legacy skill. The new paradigm is skills engineering, where the IT function or specialized departments curate a library of verified, modular skills that agents can invoke.

This shifts the burden from the individual user, who previously had to master the art of the perfect prompt, to the system architect. By templating workflows, writing style guides, and codifying business rules into reusable skills, companies can ensure consistency and reliability. The payoff is not immediate; it requires the groundwork of mapping complex business processes into machine-readable skills. Once established, these skills become a force multiplier. As the hosts noted, when a skill becomes mandatory, it eliminates hours of manual template-fitting, allowing human teams to focus on high-level analysis rather than operational drudgery.

The Recursive Advantage of Memory

A non-obvious dynamic discussed is the invisible accumulation of context. Agents like Fable 5, when given project folders and plugin access, begin to reference past successes and suggest improvements without explicit human prompting. This creates a dilemma: the more you use an agent, the better it performs, but this knowledge is often trapped within a specific, ephemeral session.

The problem I have is when I would run it again multiple times, just to double check, the only thing is my systems remember every single time I ran, I run whatever so it gets better at it. So I cannot duplicate it in another quad code or codecs because it does not have the context.

-- Andy Halliday

To capture this compounding value, practitioners must implement a dual-write regime, forcing the agent to commit its session-based learning back into the project folder. This ensures that the intelligence gained through repeated interaction remains portable and persistent, rather than vanishing when the session ends.

Key Action Items

• Audit Your IT Stack (Immediate): Evaluate whether your current infrastructure allows for role-based ID assignment for agents. If you are still relying on human-only SSO roles, you are handicapping your agentic deployment.
• Establish a Skills Library (Next 30 Days): Stop focusing on general AI literacy. Start identifying the top three most repetitive, high-friction workflows in your department and codify them into a skill that can be shared across teams.
• Implement Dual-Write Protocols (Next Quarter): If your agents are self-improving or learning through session memory, formalize a process to export that context into your version-controlled project folders. This turns hidden learning into a durable asset.
• Shift from Prompting to Templating (Ongoing): Stop training staff on how to talk to chatbots. Train them on how to build, test, and maintain the skill templates that your agents will eventually call upon.
• Adopt a Clean-Slate Pilot (6-12 Months): Identify one non-critical business process and rebuild it from the ground up assuming agentic autonomy. The discomfort of bypassing legacy IT constraints here will provide the blueprint for your long-term infrastructure.