Home / Vault & Second Brain / I Connected 314 Tools to My AI Agent. Most of Them Were Dead Weight.

I Connected 314 Tools to My AI Agent. Most of Them Were Dead Weight.

I Connected 314 Tools to My AI Agent. Most of Them Were Dead Weight.

I Connected 314 Tools to My AI Agent. Most of Them Were Dead Weight.

Last week I ran an audit on my own AI setup and the number stopped me cold: my main agent profile had 314 tools available across 11 MCP servers. Home assistant alone contributed 82. Proxmox 71. SSH manager 37. I had connected everything I could think of, the way I used to install every browser extension that looked useful. And then I never touched most of them again.

It started with a video about the ten AI skills that matter more than prompting. One line hit me directly: “Every MCP server you connect is this decade’s install-every-browser-extension mistake.” I laughed, then I checked my own config. The laugh died quickly.

The Numbers Were Worse Than I Thought

Anthropic’s research says model performance starts to degrade after about 30 to 50 available tools. OpenAI recommends fewer than 20 per turn. A study of 110 agents with 584 tools found routing accuracy dropped 16 to 23 points as the catalog grew. My default profile had 314 tools. I was miles past the point where the research says things start to break.

But here’s the nuance I found when I dug deeper: my agent uses deferred tool loading. Those 314 tools aren’t all sitting in the context window at once — they’re loaded on demand from a catalog. That softens the blow considerably. The context isn’t flooded with 314 tool definitions. Still, the catalog size matters for three reasons: the model has to pick the right tool from a bigger list, discovery costs tokens and time, and every connected server is an attack surface.


The Real Lesson: Most Tools Are Dead Weight

Then I checked which tools I actually used. One of my profiles had 104 MCP tools available and had used only 26 of them — 25 percent. The other 77 tools, 74 percent, were never called once. They just sat in the catalog, inflating the list and adding routing risk without ever doing a thing.

That’s the pattern I want to warn you about. It’s not that connecting a server is wrong. It’s that connecting everything without ever pruning is how you end up with a bloated, slower, less reliable agent. The fix isn’t dramatic. It’s segmenting servers per profile, loading tools on demand, and periodically asking: which of these did I actually use this month?

What I’m Doing About It

I’m not ripping out all my servers — that would be overcorrecting. But I’m being deliberate now. I’m considering which MCP servers each profile really needs, instead of giving every profile everything. And I’m going to monitor routing accuracy, because if vague requests start picking the wrong tools, tool surface is the first suspect.

Here’s the question I want to leave you with: how many tools does your AI agent actually have, and how many of them have you used in the last month? If you don’t know the answer, you might be carrying more dead weight than you think.

Wat vond je van dit bericht?

Leave a Reply

Your email address will not be published. Required fields are marked *