The Model Is No Longer the Moat

Artificial intelligence has spent the past few years looking in the wrong place for its centre of gravity. The conversation has been dominated by models, benchmark scores, context windows, parameter counts, training compute and the recurring question of which company possesses the most capable large language model. That emphasis made sense when access to frontier systems was scarce and when capability differences were large enough to create a genuine advantage. It makes less sense as the field matures. Organisations now have more choices, open alternatives continue to improve, switching costs are gradually falling and the gap between leading systems is narrowing across many practical tasks. The model remains important, but it is no longer sufficient to create a durable moat.

The Model Commoditises

An enterprise can choose among models from OpenAI, Anthropic, Google, Meta, DeepSeek, Alibaba and other providers. It can use one model for difficult reasoning, another for coding, a smaller model for classification and a local model for confidential workloads. That change weakens the model as the main source of competitive advantage. If the intelligence layer can be substituted, it begins to resemble infrastructure. Processors remained critical after personal computing became widespread, but most companies did not build lasting competitive advantage merely by selecting a particular processor. Cloud infrastructure became essential, but few organisations differentiated themselves simply because they ran on one particular cloud -server architecture. Artificial intelligence may be moving in the same direction.  That value is increasingly sitting in what can be described as the Harness around the model. The harness is the architecture that determines what the model can remember, what tools it can use, what data it may access and what actions it is allowed to take. It includes retrieval, memory, workflow logic, databases, evaluation, monitoring, security controls, interfaces and increasingly other models and agents.

Open models accelerate this transition because they reduce dependence. They allow organisations to experiment, adapt models for particular domains, run them locally and design systems that are less tightly tied to one commercial ecosystem. Their importance goes beyond the ideological debate between open and proprietary AI. Open models do not need to become the best models in the world to reshape the market. They simply need to become good enough for a large number of economically valuable tasks. Most enterprise workloads do not require the maximum possible intelligence available from a frontier system. A model used to classify service tickets, extract information from invoices, summarise structured reports or monitor a defined workflow may not require the same reasoning capability as a model used for scientific discovery or complex software development. A smaller or open model may therefore deliver acceptable performance at lower cost, with greater privacy and more control over deployment. This produces a more plural landscape.

LLMs Lose Centrality

This is where the idea of the fall of the large language model needs precision. Large language models are not disappearing and their capabilities are not declining. What is changing is their architectural status. The LLM is moving from the centre of the system towards becoming one component inside a broader network of specialised capabilities. The next shift follows naturally from this architecture. Artificial intelligence is moving from systems that answer questions towards systems that perform work. A conventional LLM waits for the user to provide a specific instruction. An agent can be given a broader objective and can decide what information it needs, which tools to use, what sequence of actions to follow and how to respond when the environment changes. This becomes far more significant when several agents begin working together. Consider a procurement process inside a large company. One agent could identify requirements, another could compare suppliers, another could evaluate financial impact, another could check regulatory constraints and another could examine contractual language. A human might approve the final transaction, but much of the intermediate coordination could occur among specialised agents.

In such a system, intelligence no longer resides in one place. It is distributed across several components that each contribute to the final outcome. The quality of the result depends not only on the capability of the individual agents but also on how effectively they coordinate. This is why multi agent systems may matter more than small improvements in benchmark performance. The unit of intelligence begins to move from the model to the network.

MCP Reaches Its Limit

The Model Context Protocol has played an important role in this transition because it created a more standard way for AI systems to connect to tools, applications and data. Instead of building a different integration for every model and every service, developers could use a common protocol to expose capabilities to AI systems. That was an important architectural step. However, MCP should not be mistaken for the complete architecture of agentic AI. Its primary value lies in enabling an AI system to interact with resources. An agent may need to access a database, use a calculator, retrieve a document or invoke an external service. MCP can provide a standard mechanism for that relationship.

The limitation appears when an agent needs to interact not with a tool but with another autonomous agent. A tool is generally passive and waits to be called. Another agent may have its own goals, permissions, state, capabilities and reasoning process. Communicating with that agent requires a richer form of interaction. The problem is no longer simply how to invoke a capability. It becomes how autonomous systems discover one another, describe what they can do, delegate work and exchange results. This is why it is more accurate to say that MCP has reached its architectural boundary rather than that it has died. It remains valuable, but it solves only one part of the emerging agentic stack.

A2A Extends the Network

Agent2Agent, or A2A, addresses the next layer by enabling agents to communicate with one another across different systems, frameworks and organisations. The key idea is that one agent should not need access to the internal reasoning of another agent in order to work with it. It needs to know what the other agent can do, how to communicate with it and how to interpret the result. This begins to resemble the way organisations already interact. A company does not need to understand every internal process of a supplier before placing an order. It needs a recognised interface, clear expectations and sufficient trust. Agentic systems may increasingly work in the same way.

A2A therefore extends the architecture rather than replacing MCP. One protocol helps connect agents to capabilities, while the other enables agents to coordinate with one another. The harness sits above these layers and decides how the broader system should behave. The emerging architecture begins to resemble a protocol stack for machine intelligence. Models provide reasoning. MCP provides access to resources. A2A provides communication between agents. The harness provides orchestration. Human governance determines what the entire system is permitted to do.

Ethics Moves to the System

This architectural transition has major consequences for the Ethics of artificial intelligence. Much of the existing discussion around AI ethics has focused on model behaviour. We ask whether a model is biased, whether it hallucinates, whether training data was obtained fairly, whether harmful content can be generated and whether the system is sufficiently explainable. These questions remain important, but agentic systems create a different category of ethical problem. An inaccurate chatbot can provide bad advice. An inaccurate agent with permission to move money, modify a medical record, approve a supplier, change a production system or access confidential information can create much more serious consequences. Once AI moves from language generation into action, authority becomes an ethical issue.

The relevant questions therefore become broader. Who authorised the agent to act? What information was it allowed to access? Could it delegate authority to another agent? Could that second agent delegate it again? What happens when several agents collectively contribute to an outcome that none of them independently intended? Can the sequence of decisions be reconstructed afterwards? Who is accountable when the final action emerges from a network rather than a single model? These are no longer only questions about alignment. They are questions about legitimacy, responsibility and institutional authority.

Identity becomes one of the most important foundations of agentic AI because a system cannot govern what it cannot properly identify. Before an autonomous agent is allowed to act, another system needs to know what that agent is, which organisation it represents, who authorised it, which permissions it possesses and how long those permissions remain valid. This distinction between capability and authority is crucial. A highly capable agent may still be safe if its permissions are narrow and its actions are tightly controlled. A less capable agent may become dangerous if it has unrestricted access to sensitive systems. Ethical risk therefore depends not only on how intelligent the system is but also on what the system is empowered to do. The next phase of AI governance will therefore have to focus much more heavily on identification, authorisation, revocation, auditing and delegation.

Ethics Becomes Infrastructure

The most important consequence may be that ethics becomes part of technical architecture rather than a policy added after deployment. Identity management becomes an ethical mechanism because it determines who can act. Permission controls become ethical mechanisms because they determine what actions are legitimate. Audit trails become ethical mechanisms because responsibility requires evidence. Human approval points become ethical mechanisms because some decisions should not be fully delegated. Data boundaries also become ethical architecture. An agent may technically be able to share information with another agent, but technical possibility does not establish moral permission. Observability becomes essential because an organisation cannot govern a network whose behaviour cannot later be reconstructed. The ability to stop an autonomous workflow, revoke credentials or prevent further delegation is therefore not merely a security feature. It becomes part of responsible AI.

Beyond the Model

The first race in generative AI was about building the smartest model. The next race will be about building the most capable, reliable and governable system. The future of AI is therefore unlikely to consist of one enormous model performing every cognitive function. It is more likely to contain ecosystems of intelligence in which large models, small models, open models, proprietary systems, deterministic software, specialised agents and human decision makers operate together. This is why the model is no longer the moat. The model may become one of the easiest parts of the architecture to replace. What will remain difficult to replace is the knowledge surrounding it, the workflows connecting it, the trust embedded in it and the governance determining what it may do.

That is where the next moat will be built and that is also where the Ethics of Artificial intelligence will increasingly reside.