Every character in that description gets injected into the model’s context. On every turn that involves tool use. Here’s what the difference looks like: Before an agent even begins solving the user’s problem, it may be processing thousands of tokens describing available resources. The larger the ecosystem becomes, the larger this overhead grows. What starts as a few lightweight integrations can evolve into an enormous context payload that accompanies every engagement. Most of that information may never be relevant to the user’s request. Yet the model still has to read it.

The Hidden Cost Nobody Talks About

Token bloat is often discussed as a technical issue, but its business effect is equally important. Large prompts consume more tokens. More tokens mean higher inference costs. Higher inference costs translate directly into larger infrastructure budgets. However, cost is only one piece of the puzzle.

Latency increases because the model must process larger inputs before generating an answer. User experience suffers because responses feel slower. In real-time business workflows, even a few additional seconds can greatly affect productivity and satisfaction. There is also an accuracy challenge. Contrary to popular belief, providing a model with excessive context does not guarantee better reasoning. Models can become distracted by irrelevant information, lose focus on the core objective, or prioritise details that have little value to the task at hand.

More information does not always create more intelligence. Sometimes it creates noise.

Context is Becoming the New Technical Debt

Software engineers have long understood technical debt. Small shortcuts accumulate over time until systems become difficult to maintain. Token bloat is rapidly becoming the AI equivalent. Teams often add new MCP servers with good intentions. Every new business application, database, knowledge repository, and service appears valuable. Nobody wants to remove functionality. Nobody wants to limit access. As a result, the context footprint grows continuously. Months later, organisations discover their agents are carrying around massive amounts of metadata, tool definitions, and operational descriptions that have accumulated through incremental decisions.

The challenge is rarely caused by one big mistake. It is usually the result of hundreds of small additions that appeared reasonable at the time.

Why “Everything Everywhere” Is the Wrong Strategy

One of the most common design mistakes in agent architecture is providing universal access to every available tool. The thinking seems logical: if an agent can access everything, it can solve anything. The reality is different.

Highly effective AI systems tend to be selective rather than exhaustive. They focus on delivering the right information at the right moment instead of making every possible piece of information available all the time. An employee requesting annual sales figures does not need the agent to load manufacturing telemetry schemas, HR workflows, procurement APIs, or facility management tools. Yet many implementations do exactly that. Context should be dynamically assembled based on intent rather than statically injected into every engagement. Smart filtering often delivers more value than unlimited access.

The Rise of Context Engineering

As AI systems mature, prompt engineering is giving way to something much bigger: context engineering. The challenge is no longer simply asking the model the right question. The challenge is to ensure the model receives only the information required to answer the question effectively. Organisations are increasingly realising that successful AI architectures depend on careful context management. This includes selective tool discovery, smart retrieval strategies, context compression, metadata optimisation, and dynamic loading mechanisms. Instead of sending entire catalogues of capabilities to the model, modern architectures aim to provide just-in-time context. The model doesn’t need access to everything. It needs access to the right thing.

Less Can Be More

There is a fascinating paradox in AI systems. The most sophisticated architectures are often the ones that send less information, not more.

• They remove redundancy.
• They eliminate unnecessary metadata.
• They prioritise relevance over completeness.
• They treat context windows as a precious resource rather than an unlimited storage space.

This mindset shift is critical as organisations scale their agent ecosystems. Context capacity may continue to grow, but demand for context grows even faster. Without discipline, token bloat will simply expand to consume every available resource.

Sound familiar? It is the same story we have seen with memory, storage, network bandwidth, and cloud spending.

Is There a Future for MCP?

Absolutely.

MCP arrived at the perfect time, providing a standardised way for AI agents to connect with enterprise systems, tools, and data. As organisations scale their AI capabilities, the need for interoperability and consistency will only increase. However, the future of MCP is not simply about connecting more things. It is about connecting them intelligently. The next generation of agent architectures will focus on delivering the right context, tools, and data at the right moment, rather than overwhelming models with all available information.

This is where the challenge of token bloat becomes critical. The value of MCP will depend not on how a great deal of context it can provide, but on how effectively that context is curated and managed.