The Subtraction Principle: Why Mature AI Agents Do Less, Not More
Vercel made its sales agent better by deleting 80% of its tools, not adding more. Why removing capabilities — not adding them — is the real marker of agent maturity.
Vercel made its own sales agent better by deleting 80% of its tools, not adding more. That single fact cuts against most of what gets said about AI agents right now, that more integrations, more memory, more autonomy compound into better performance. Past a point, they stop compounding. They start eroding trust.
Why would removing tools make an agent better, not worse?
Vercel didn't start there. The team studied one of its strongest reps closely enough to see the actual shape of the work, which messages were real leads, which were support questions in disguise, what the rep researched before replying, where judgment had to stay human. They built the agent around that observed workflow: filter inbound messages, qualify leads, research companies, draft responses, route support away from sales, then hand everything to a human for review.
That's already a solid build story. The more interesting part is what happened next. The team kept adding tools, memory, and integrations the way most of us do once something starts working, and the agent got harder to trust, not more capable. It only improved again once they cut 80% of what they'd added. The beginner instinct is to keep adding. The instinct that actually holds up over time is asking what should come out.
What's really being maintained, the model, or the setup around it?
Most of the maintenance burden isn't the model. It's everything around it: what the agent is allowed to read, what it can touch, what it's permitted to change, what proof it has to bring back before anyone trusts the output. That setup has to move as the model underneath it changes, and that's a genuinely new problem. Traditional software breaks when it gets worse. Agents can break when the model gets better.
A permission that was harmless for a weaker, more cautious model can become too broad once a stronger model starts acting on it confidently. A restriction built to contain a clumsy model can quietly throttle a capable one. Neither failure is loud. The agent keeps producing work either way, it just becomes wrong work, or overreaching work, without announcing which.
Isn't "add more capability" the whole pitch of this industry?
It's the pitch, but it isn't what's actually surviving contact with production. Gartner projects that more than 40% of agentic AI projects will be shut down by the end of 2027, and the reasons cited are cost, unclear business value, and inadequate risk controls, not that the underlying technology doesn't work. Scope crept until nobody could say precisely what the agent was for or where its edges were.
The discipline shows up earlier than most people expect, too. One operator building a personal automation tool gave an AI full access to research his business context, except one Slack channel, walled off on purpose before the tool ever ran. Not because the model couldn't handle it, but because some information simply had no business being in an agent's reach. That's subtraction as a design decision made up front, not a cleanup performed after something goes wrong.
| Beginner instinct | Maintenance instinct |
|---|---|
| Add tools until it can do everything | Ask what it should stop being able to touch |
| Broad access, "just in case" | Access scoped to what's actually used |
| More memory, more context | Question whether that memory has gone stale |
| Success measured by feature count | Success measured by output someone actually acts on |
A couple of questions this tends to raise
Does this mean agents should start narrow and stay that way? Not quite. Vercel's agent grew broad first, built around real observed work. Subtraction was a second phase, not a starting posture, you generally need to see what gets used before you know what to cut.
How do you actually decide what to remove? Look at what the agent is reading and whether that source is still current, what it's allowed to touch versus what it merely could touch, whether its job has quietly expanded without anyone deciding that on purpose, and whether the thing it produces gets acted on or just piles up unread.
Reliability, on this evidence, isn't a feature you bolt on. It's what's left over after everything that wasn't earning its place has been taken back out. Worth a look at your own setup before the next tool gets added.
Sources
- Vercel (n.d.) Reportedly improved agent reliability by removing 80% of its tools. Referenced in source material; original company disclosure not independently reviewed by Markedeen.
- Gartner (2026) More than 40% of agentic AI projects will be discontinued by the end of 2027. Referenced in source material; original report not specified.
Want a system like this in your business?
We build the automation behind everything you just read.