Functional AGI is here: ComputerUse & Coding agents can collaborate using MCP achieved near-AGI autonomy enough for enterprise automation at scale!

When I asked this morning Microsoft (Powerpoint) copilot to read my managed service solution pitch deck, pick up service catalogue from slide 7 and generate a potential skill matrix table for the whole catalogue on an excel sheet, it provided me the sheet in matter of 4 secs.

What exactly happened here ? Less magic, little bit of reasoning, little bit of general knowledge and lots of python coding!

How the Agent Planned & Executed to achieve this goal :

The Office Agent in Microsoft 365 Copilot is built upon an open-source stack, primarily leveraging Anthropic's Claude code and OpenAI models. To generate, manipulate, and analyze files (such as PPTX, Excel, and Word), it employs a Python-based code interpreter in a secure, sandboxed environment (on company's private azure cloud), running python code generated by (lets assume) ClaudeCode model.

1. Environment Setup: The agent ensures python-pptx and pandas are available in its runtime environment.
2. Contextual Extraction: It doesn't just "read" everything; it specifically indexes the slides[6] object. If the service catalogue is in a table, the agent iterates through shape.table.rows. If it's a list, it iterates through text_frame.paragraphs.
3. Heuristic Reasoning: The most "AGI-like" part of this code is step 4. The agent must infer what a "skill matrix" looks like. It defines columns like "Technical Architect" or "Project Manager" based on its internal knowledge of managed services, even if those words weren't on Slide 7.
4. State Persistence: The generated .xlsx file is stored in the agent's temporary session storage. It then triggers a UI "artifact" (the Excel download button) for you.

This is a classic simple example of 'Code is the way to achieve any goal' by AI agents. Artificial General Intelligence (AGI) in coding has not been universally achieved yet, while advanced AI models like Claude Sonnet 4.5 or Opus 4.5 have demonstrated human-level, and sometimes superhuman, proficiency in generating, debugging, and understanding complex code, AI has reached over 71% on benchmarks like SWE-bench (software engineering tasks), a massive jump from just 4–5% in early 2024. But to generate codes to automate processes and move data to drive enterprise agentic automation at scale we may not need further improvement, this is not coding to build complex application, this is about coding to automate, we may have achieved the fundamental power behind 'Functional AGI'.

The quest for Artificial General Intelligence (AGI) is often portrayed as a race toward a digital consciousness. In the context of Industrial Revolution 5.0, the most relevant definition of AGI is not philosophical; it is the ability to fully autonomously execute complex tasks over extended periods to deliver measurable economic value.

But while the industry focuses on replicating human biology through massive compute, Functional AGI has already arrived in the context of driving autonomous enterprise automation at unprecedented scale, accuracy & predictability. It is infact passing through, there is a visible AI overhang, where capability remain largely unused by global enterprises.

Bridging the Agent (LLM) Intelligence Gap

Human intelligence is defined by dynamic, experiential learning and adaptability. In contrast, standard Large Language Models (LLMs) rely on pattern recognition within static training data. The "intelligence gap" has traditionally been the inability of AI to handle real-time context. The steady advancement of Model Context Protocol (MCP) servers solves this, acting as a nervous system that connects LLMs to external resources, knowledge graphs, and private enterprise data, at the right time with the right amount of context.

Model Context Protocol (MCP) has evolved from a niche, open-source initiative aimed at solving AI-tool integration fragmentation into a rapidly adopted industry standard, growing from its inception in November 2024 to a "de-facto" standard by late 2025. It has grown from a simple "connector" to an "agentic middleware" that enables AI systems to securely access data and tools, with thousands of servers developed for tasks ranging from file system access to web search and database interaction.  Anthropic released MCP as an open-source standard to address the AI tool problem, where each AI model required custom integration for every external tool. Later Anthropic donated the MCP to the Agentic AI Foundation (AAIF) under the Linux Foundation in December 2025 to foster further open-source development. 

• Shift from "Model-Centric" to "Context-Centric": As LLM intelligence hits limits, the focus has shifted to providing better context. MCPs provide a, "structured way to describe" what a system does, which is more effective than raw, unorganized context. The future of MCP is expected to move toward "multi-model orchestration," where MCP serves as the glue for connecting different AI models (LLMs, vision, voice) with diverse tools, and the development of a "unified MCP marketplace" for easy discovery and deployment. 

The New Agentic Workforce

We are seeing the emergence of two specialized agent archetypes that, when combined and empowered with the advanced capabilities of MCP, create "near AGI" capabilities across any application stack or operating system or business function or business process.

To achieve autonomous, enterprise-grade Agentic AI, we must move beyond traditional task automation toward reimagined workflows (junking old processes designed for human workers) —where dataflows (inputs from customer, suppliers, functions, stakeholders) merge with decision flows (company policies, rules) to form truly autonomous business processes. In this model, agents replicate two essential human-like skills supported by a persistent "nervous system":

• Application Orchestration (CUAs): Computer Use Agents navigate enterprise UIs just as humans do, performing discrete transactions within ERPs or CRM systems. In-applications steps.
• Intelligent Data Intermediation (Dynamic Coding): Coding agents generate on-the-fly logic to process documents, apply business rules, and move data across the enterprise boundary or between siloed applications. Between applications steps.
• The Foundation (MCP): State-of-the-art Model Context Protocol servers provide the persistent memory and "tribal knowledge" required for complex, autonomous agent-to-agent interactions. Dynamic context in a rail.

Real-World Example: Autonomous PO Processing

In a supply chain context, an Autonomous PO Agent identifies a stock shortage via a Supply Chain MCP. A CUA logs into the supplier portal to check live availability, while a Dynamic Coding Agent extracts pricing from an incoming (outlook mail) quote PDF, applies the company’s "Best Price Policy" (Decision Flow), and autonomously generates (by a CUA agent) a Purchase Order in the ERP. If the price exceeds a threshold, the agent initiates a chat (API call through MCP) with the Procurement Manager—summarizing the conflict and providing the essential data needed for a final human-in-the-loop sign-off.

Functional AGI is Here

When a multi-agent workflow can break down a multi-hour business process, leverage a CUA for UI interaction, and use a Dynamic Coding Agent for data synthesis—all supported by purpose-built MCP servers—the goal of AGI has been met. Whether it is an off-the-shelf MCP for a standard ERP or a custom-built one for a proprietary app, the infrastructure for autonomous value creation is now in place.

We do not need to wait for a model that "thinks" like us. If it can autonomously navigate the complexity of modern business to deliver outcomes, the threshold has been crossed. Functional AGI is no longer a future milestone; it is a present reality passing right before us. A significant economic opportunity is on the table, are we ready with the right leadership vision, enterprise agentic infrastructure, placing bets on AI disruptions and above all, reimagining the future operate!

** All views here are purely personal and not of EY's.