Your AI Agent Should Have Fewer Skills, Not More

Every AI agency you talk to will brag about the number of tools their agents can use. "Our agent has 47 skills." "Ours has over 100 integrations." "Our platform supports 200+ actions."

That number is not a feature. It's a warning label.

#What actually happens when you add the 30th tool

Agents pick tools based on their descriptions. When your agent has three tools, the choice is easy. Read one description, read another, pick the one that fits. When it has thirty, the choice is not easy. It's an averaging problem across thirty descriptions, and the model resolves it with a shrug. It picks the tool whose description sounds most like the current task, which is often the newest one, which is often the wrong one.

We watch this happen in production. An agent has a search tool, a database tool, and a memory tool. If they overlap in description ("search for information"), the agent will pick whichever is at the top of the list, or the last one it used, or the shortest name. Which of those it picks on any given day is not signal. It's noise.

Add ten more tools with similar surfaces and your agent stops being deterministic. It stops being debuggable. It stops being an agent and starts being a slot machine.

#The math is against you

Every tool has three things attached to it:

• A description, which lives in your prompt on every call.
• A test surface, which you either have or don't.
• A failure mode, which you'll discover in production.

Ten tools with mediocre descriptions is a bigger prompt tax than three tools with sharp descriptions. Fifty tools is a full page of prompt bloat before you've written a single instruction.

The test surface is where it really breaks. If you have five tools and you actually test each one against the agent's decision-making, you have five tests. If you have fifty, you have almost certainly tested none of them properly. You are one edge case away from the wrong tool firing at the wrong moment on the wrong customer.

#What we actually ship

Our production agents have between five and twelve tools each. Not because we're minimalists for the sake of it. Because that's the count where the agent still picks correctly.

• The CEO agent has around ten.
• The Growth agent has closer to eight.
• The specialist agents (Web Designer, Creative Producer) have their own tighter sets, because they do one thing.

When we've been asked to add a tool, our first question is usually: which existing one does this replace? If the answer is "none, it's additive," we push back. Not always successfully. But we push back.

#What to do this week

If you are buying an AI agent from a vendor, do this:

• Ask how many tools the agent has today.
• Ask what tools it uses most often. If they can't tell you, they aren't watching.
• Ask which tools they've removed in the last six months. If the answer is "we don't remove tools, we only add them," walk away.

If you are running your own agent:

• Print the list of tools your agent has right now.
• Group them by intent. If two tools cover the same intent, one of them is dead weight.
• Delete the one you use less. Rerun your evaluation suite. If nothing broke, you found dead weight. There will be more.

#The unpopular part

Fewer skills means saying no to internal requests. It means telling a stakeholder that no, you will not add "one more small integration." It means the agent's capability list is shorter than a competitor's, which will read as a weakness in a sales pitch even when it's a strength in production.

We accept that trade because we would rather have an agent that picks the right tool 98% of the time from a list of ten than one that picks the right tool 60% of the time from a list of fifty. The customer feels the difference. The vendor feels good about the second one, right up until the churn.

If you want us to audit your agent's tool set and cut the dead weight, start here.