Plugging an off-the-shelf MCP server, with an upstream defined schema directly into an agent is a textbook example of the Conformist pattern. You are injecting an externally defined interface straight into the agent’s system prompt – the absolute core of its bounded context.

When tool definitions are exposed naively, an upstream API change can instantly alter the tool schemas passed to your agent. Your application code won't catch this breaking change. Instead, it will pass silently into the agent's context window until the LLM unexpectedly hallucinates or fails.

To be fair, the culprit isn't MCP's underlying protocol, but the architectural malpractice it facilitates. It makes it dangerously simple to skip proper system design and let raw, uncurated upstream data structures dictate your agent's cognitive model.

Beyond architectural fragility, this conformist approach introduces immediate operational inefficiencies. An off-the-shelf Jira MCP server typically exposes dozens of generic tools. Your specific coding agent might only ever need two or three of them. Yet, the entire uncurated toolset remains in the payload, unnecessarily bloating the agent’s context window and driving up token costs.

You can mitigate this, of course, by filtering the list of allowed tools. But how do you enforce your internal project conventions? For example, if you want the agent to create a new user story, how do you ensure it follows your specific team template and engineering criteria?

Today, the industry's default band-aid is to cram these business rules into the system prompt, either directly into an agent.md file or encoded as a prompt-based "agent skill."

While this might work for a local coding assistant – where a developer is sitting right there, carefully observing every output – the calculation completely changes for an enterprise agent. When you scale to autonomous, production-grade workflows, look at what you’ve actually built: your prompt has become your anti-corruption layer.

And let’s be honest: in an enterprise environment, do you really want your critical architectural boundaries to be governed by a piece of probabilistic prose?

The limitations of the Conformist pattern become especially clear when we move from single-tool integrations to complex enterprise workflows.