Tech insights: Secrets of AI Agents

What happened?

AI agents are not new. Software agents, autonomous units of software that respond to requests from users or other programs, were first conceptualised in the 1950s and became mainstream in the 1980s through patterns such as distributed systems and microservice architectures. Familiar examples followed, from Microsoft’s Clippy in 1996 to Amazon’s Alexa Skills Kit in 2015.

Today’s generative AI community has co-opted the “agentic” term to mean an end-to-end application ecosystem built on large language models, which has caused confusion about what is actually running underneath the label.

An agentic framework can be broken down into a small number of roles: a Client that makes a request; a Parser, Reasoner and Orchestrator that interpret it, decide how to respond, and route sub-tasks; and one or more Servers, the “agents” themselves, that register what they can do and respond to requests.

Why does it matter?

The ease of entry into agentic AI is pushing companies to adopt it without looking closely enough at the AI, or other methods, being used underneath the agentic covers.

Before large language models, communications between software components relied on formal, fragmented standards such as SOAP, CORBA and OAS. Anthropic’s open-sourced MCP now provides a shared way for agents to register skills and communicate using natural language, and Google’s emerging A2A protocol looks set to standardise how agentic servers talk to one another.

Who is affected?

Wealth management and compliance teams evaluating or procuring agentic AI tools, and vendors building agentic products, who need a shared understanding of what is genuinely AI-driven inside these systems.

Key risks

• Assuming an “agentic” label automatically means embedded AI, when the underlying logic may be predefined, rule-based software.
• Not understanding which components, the parser, reasoner or orchestrator, are AI-driven versus rule-based.
• Confusion caused by inconsistent industry use of the terms “agent” and “agentic”.

Actions to take

1. When evaluating an agentic AI tool, ask which components are genuinely AI-driven versus rule-based.
2. Check whether the tool follows an interoperability standard such as MCP, or the emerging A2A protocol.
3. Look past marketing language to understand the actual communications contract and skills each agent registers.

Wider implications

As large language models increasingly act as the shared communications protocol between software components, and standards such as MCP and A2A begin to formalise how agents register skills and talk to each other, understanding these roles will matter more for firms building on or procuring agentic tools.