Catching the silent failures of MCP servers

Why this conversation

Building an MCP server has become easy; knowing what agents do with it once it is out is not. Kashish Hora works at AgentCat (“first of all, we are a bunch of founders and cat lovers and dog lovers”), which helps the companies that build MCP servers see how agents use them and where they get stuck. The host talked to him in San Jose on Friday 23 October, the second day of AGNTCon + MCPCon North America. Everyone else the show met that Friday is in AGNTCon + MCPCon North America 2026: day two.

Silent failures

“Building an MCP server is really easy nowadays. There’s many tools to help you get started.” The hard part is what comes after the release. Logs and traces catch the errors that are easy to catch, such as an authentication failure. What they miss, Kashish says, are silent failures: an agent that tries to do something for the user and cannot, because the server lacks the function or it is hard to find, and gives up, abandons the session early, or calls the wrong tool and gets confused. For the user, it looks as if the tool does not support what they asked.

AgentCat’s SDK is one line to install, in the major languages MCP servers are written in (“TypeScript, Go, Python, Ruby, whatever it might be”), and it instruments the server automatically, almost like middleware. The agent can then report back what it could not do, and the developer collects that feedback over time to improve the server. The data can also go to any OpenTelemetry collector. “So it’s not built on top of OpenTelemetry, it exports OpenTelemetry”, the host checked. “Exactly.” The goal is not to replace the analytics and observability already in place: “Our goal is to give you an entirely new way to understand how agents are experiencing your MCP server.” Nor is it a gateway; it lives on the MCP server itself.

The agent experience

For two decades, companies built products around how people use them: websites, web apps, forms. Kashish expects the most successful companies of the next two decades to be the ones that prioritize how agents experience their product, because agents will be the biggest users of software. Today that mostly goes through MCP, “but you’re also seeing all these agents try workarounds like just opening your website and trying to click on buttons in a way that was meant for humans”. AgentCat starts with MCP and is working on WebMCP, plain browser use, and standards such as A2A and UCP. The host, who had been at Nordic APIs, raised the problem that agents do not know a user’s workflow and try every path. “It’s a different way of thinking about software development, and for a lot of organizations, it’s a new muscle”, Kashish answered. He also sees new roles appear for this, such as head of agent experience. Webflow is one of the customers it has helped as its MCP server grew.

Getting started

AgentCat has “a website and a product that’s fully self-serve and free to try and free to use. All of our SDKs are open source. You can get started at agentcat.com.”

When it comes out

The episode is edited and in the release order. Like every episode it goes out on YouTube first, then on Apple Podcasts, Spotify and Amazon Music.