The rush to build AI agents for customer service is now well under way. Contact centers, CRMs, and workflow tools all promise “agents” that can understand requests and take action. As Ian Jacobs argued in his recent Opus Research post on the need for an AI agent control plane, the real challenge is not getting any one agent to work. It’s managing the sprawl: many agents, on many platforms, each with their own rules, tools, and memories of the customer. 

Ian’s control plane is an enterprise layer that sits above channels and applications. It keeps track of the customer’s journey, enforces policy, governs knowledge, and evaluates whether any of these agents are actually improving outcomes. Think of it as the governance and consistency layer for agentic CX. 

If we dive one level below the enterprise control plane, we come to the execution layer, which controls how agents do actual work. Here, a new idea from Anthropic is gaining traction. That idea involves the creation of a new “agent skill” layer. 

What is the skill layer? 

Anthropic recently spelled out this new Agent Skill layer in detail.  Inside any given platform, a skill is a small, named capability that an AI agent can invoke when needed. It might: 

• Recognize what a customer is trying to accomplish 
• Rephrase a question so a search system can answer it 
• Walk through a standard process such as updating an address, checking eligibility, or issuing a refund within policy 

These skills sit between the general-purpose model and the underlying systems. The model still handles listening, reasoning, and responding. But when it comes to doing real work, it calls skills that encode how your business actually operates. 

Inside a skill, there may be tools and APIs that behave deterministically, for example to check a balance, change a booking, log a note in CRM, and send a confirmation. Those steps should not be left to probabilistic generation. They are implemented as code. Skills tie those tools together with instructions and guardrails that the model can understand. 

Most implementations also introduce an orchestrator agent. This is a kind of supervisor that receives the customer request, decides what the intent is, and then sequences the necessary skills and tools. One interaction might involve a “clarify intent” skill, a “summarize for agent” skill, and a deterministic tool that changes an order, all chained together. 

So, if the control plane defines the shared rules and context for all agents, the skill layer is where those rules become concrete, reusable tasks that any agent can perform inside a given platform. 

Why skills matter for CX 

Looked at through a CX lens, the skill layer solves several practical problems that tripped up earlier generations of bots. 

First, skills make behavior more reliable and auditable. A general model does not know your refund policy or your identity-verification steps until you tell it. If you cram everything into one enormous prompt, behavior becomes fragile, and every change becomes problematic. Skills let you isolate procedures into units you can inspect, test, and update without breaking everything else. 

Second, skills enable reuse across channels and agents. The same “verify identity” or “check upgrade eligibility” skill can support a self-service bot, an outbound agent that does proactive outreach, and an agent-assist workflow in the contact center. When something changes because of a policy tweak or a new compliance requirement, you only need to update the skill once. 

Third, skills clarify the split between conversation and procedure. The model can be flexible and human-like in how it talks. The skill is strict about the steps that must be followed and the tools that must be called. That division becomes more important as enterprises start using a control plane to enforce policy across multiple platforms. The control plane can say “identity must be verified before this action,” but the skill is where the concrete steps live. 

Finally, skills provide a natural place to capture expert knowledge. Every contact center has experienced agents who “just know” how to handle edge cases. A skill gives you a place to write that down in a way that an AI agent and new human agents can reliably use. 

How skills fit under an AI agent control plane 

In a multi-vendor future, most large organizations will not standardize on a single agent platform. They will have CCaaS providers, CRM vendors, specialized workflow tools, and possibly their own in-house frameworks. Ian’s control plane is about creating a shared layer of rules, identity, knowledge governance, and evaluation that spans that mix.  

Within that world, the skill layer plays a vital role. Skills act as the execution fabric for control-plane decisions. If the control plane decides that a certain journey state, identity, and consent level allow a self-service agent to change a subscription, the actual steps are implemented as skills inside one or more platforms. 

Skills also provide standardized building blocks that can plug into cross-platform rules. In addition, they become the main surface where continuous evaluation feeds back into execution. If the control plane’s telemetry shows that “simple billing question” journeys are failing too often, you may discover that a particular “explain my bill” skill is confusing or mis-ranking knowledge sources. That is a surgical fix. You revise the skill, not the entire environment. 

What enterprises can do now 

For enterprises, the practical opportunity is to start treating skills as long-lived assets rather than by-products of implementation projects. A sensible way to begin is to pick a handful of high-volume, high-impact journeys you would want any future control plane to govern well. Then describe the key functions of these journeys as skills.  

Those skills should be versioned, monitored, and improved based on real interactions. Over time, you will accumulate a catalog of enterprise-specific skills that embody how your best agents already handle these journeys. When a control plane arrives, either from a vendor or assembled by your own architects, it will have something concrete to govern. 

Agentic CX is not just about standing up more agents. It is about governing a growing mix of agents and making sure they execute your business the way you intend. The AI agent control plane is one answer to the governance problem. The recently defined Agent Skill layer framework could be the execution answer underneath it. 

Categories: Articles