By Michael Domanic, Section Head of AI

You’re probably paying for agents that everyone has forgotten about. 

Here’s what happened: Someone on your team built an agent a few months ago. It ran for a week or two. It didn't quite produce what they wanted. They got busy. They moved on to the next thing.

The agent didn't.

It's still running, firing on its schedule or whatever trigger it was originally set to. It’s still burning tokens. And nobody remembers it exists. 

Multiply that across an entire organization, and you’ve got a real problem. Ghost Agents are quickly becoming the number one source of wasted inference spend, and most organizations have no way to even identify them.

You can't audit what you can't see

The real issue underneath all of this is visibility. Most organizations cannot answer a basic question: What agents are currently running in our environment, who built them, and what are they doing?

Your work starts with making the agent landscape visible. And critically, that visibility has to be baked into the system. It cannot depend on people remembering to document what they built, because they won't.

Here's how we're approaching this at Section.

1. Every agent gets a 30-day expiration. This is one of the most important design decisions we've made to address the Ghost Agent problem. When someone deploys a new agent, it's set to run for 30 days maximum. Before that window closes, we audit it. Is it producing what we expected? What does it cost to run? Does it need to be adjusted to add more value? If it’s showing value, we renew it and set the next review. If nobody shows up to make the case for it, the agent dies automatically. This alone would solve the ghost agent problem for most organizations. If you do nothing else from this article, do this.

2. Agents watching agents. Nearly everyone at Section has a dedicated AI World Slack channel. An agent reviews their context, what they're working on, what they're building, what tools they're using, and then posts a daily report. These reports aren't designed for humans to read - they're structured so that other agents can subscribe to them and consume the information. We've built an ambient awareness layer across the company where the system itself knows what's being built, by whom, without anyone filling out a form.

When someone leaves, I don't have to go on a scavenger hunt through their environment to figure out what they were running, because the record already exists in Slack. 

3. Automated usage auditing. We're building an agent that monitors tool usage across the org and produces a weekly report on what's running, what it's consuming, what it's producing. If an agent has been burning tokens for three weeks and nobody has interacted with its output, the agent flags that to us. 

4. Quarterly human review. I’m trying something new this quarter: blocking off a standing, two-day office hours to help Section employees audit their agents. I’ll ask employees to book 30 minutes with me, and we’ll review three questions per agent:

• What's working, and how do we amplify it? 
• What's not working, and should we kill it or suspend it? 
• What's the cost versus the value it's generating?

These sessions are deliberately framed as support, not surveillance. The goal is to help people get more value out of what they built. Nobody is getting in trouble for experimenting, but experiments that didn't work out don't get to run forever on the company card.

The panic response (please don't do this)

My biggest advice: Don’t let your leadership team panic and shut down agentic capabilities because the bill has come in. We see this every day in the enterprise. The CFO starts asking questions, and someone in the room suggests the obvious: set stricter limits on who can build agents. 

If you've been reading this newsletter, you know where I stand on this. Restricting access kills the experimentation culture you need. You’ve spent months telling people to build things and enabling them to do so. The moment you put a gate in front of that, you're punishing the builders for what should be considered a management problem.

You don't need approval gates. You don't need to slow anyone down. You need a system that makes your agent landscape visible by default, automated tracking that flags what's running and what it costs, and a regular conversation with your people about what's worth keeping alive.

Build fast. Audit regularly. Kill the ghosts.

See you next week,

Michael
Your fellow Head of AI

‍