The pattern repeats across large companies. A year of enthusiasm produced agents everywhere. One team built an agent to summarize tickets. Another built a near-identical one. A third wired an agent into a workflow, shipped it, and moved on, and now nobody owns it. Each agent calls a model, so each one has a bill and a data path, and most of them were never inventoried.

This is shadow IT with a meter running. The unmanaged spreadsheets and rogue apps of the last decade did not bill you per use. Agents do. Every redundant agent is a recurring token cost and a governance exposure, and the more of them you have, the more of both you carry.

I run cost and governance on enterprise Salesforce programs, where unmanaged anything turns into revenue leakage. On the T-Mobile CPQ and CLM work I lead through Mphasis, the savings come from controlling sprawl: removing duplicate processes, retiring what does not pay for itself, and keeping the few things that move a number. That same discipline applies directly to agents. Here is the detox I would run.

Step 1: Inventory every agent, including the ones nobody claims

You cannot manage what you have not counted. Start with a full inventory of every agent calling a model in your environment.

Pull it from where the money and the calls are. API billing dashboards. Model gateway logs. The list of service accounts hitting your model endpoints. Each of those calls traces back to an agent, and the orphans, the agents with no clear owner, are exactly the ones the audit exists to find.

For each agent, record: what it does, who owns it, what it calls, how often, and what it costs per month. If you cannot fill in the owner, flag it. Ownerless agents are the first cleanup target.

Step 2: Map overlap

Group the agents by job. You will find clusters doing the same work: three summarizers, four classifiers, two that draft the same kind of reply.

Overlap is where the easy savings are. When five agents do one job, you are paying five token bills and maintaining five data paths for one outcome. Consolidating to one removes four bills and four exposures with no loss of function.

Step 3: Score each agent on cost versus value

This is the core of the detox. For every agent, put two numbers side by side.

Cost is straightforward: monthly token spend, plus the maintenance and the governance overhead of keeping it. The token number comes from your billing data and is exact.

Value is harder and worth the effort. What measurable outcome does this agent produce? Time saved, revenue influenced, errors caught, work avoided. If a team cannot name a measurable outcome, that is your answer. An agent with a real bill and no nameable value is a pure cost.

Plot them. Agents with high cost and low value are the first to retire. Agents with low cost and high value are safe. The middle gets a closer look.

Step 4: Check the governance exposure on each one

Cost is half the risk. The other half is what the agent can see and do.

For each agent, ask: what data does it read? What can it write or change? Does it have access it does not need for its job? An agent with broad permissions and a vague purpose is a security problem waiting to surface, regardless of its cost.

I treat this the way I treat a quote that can apply an unapproved discount. The exposure is not theoretical. An over-permissioned agent is a path to data you did not mean to expose, and every extra agent is another such path. Retiring redundant agents shrinks the attack surface as a side effect, which is reason enough on its own.

Step 5: Retire, consolidate, or keep

Now act on the scores. Three buckets:

• Retire: high cost, low value, or ownerless, or redundant with a better agent. Turn it off. Watch for a week. If nothing breaks, decommission it and revoke its access.
• Consolidate: part of an overlap cluster. Pick the best one, route the work to it, retire the rest.
• Keep: clear owner, measurable value, scoped permissions, justified cost. Document it so it does not become next year's orphan.

Turning an agent off is reversible and informative. If a week passes and no one notices, you have your proof that it was not earning its bill.

Step 6: Put a gate in front of new agents

A one-time cleanup that is not followed by a control just refills. Sprawl came from the fact that anyone could stand up an agent with no review, so close that gap.

A lightweight gate is enough. Before a new agent goes to production, it needs a named owner, a stated measurable outcome, scoped permissions, and a cost estimate. That is four fields, not a committee. It stops the easy duplication and the ownerless deploys without slowing real work.

What the detox is worth

The return shows up in three places. A lower model bill, because you stopped paying for redundant inference. A smaller attack surface, because you removed access paths nobody was watching. And a clearer picture of which agents actually drive an outcome, which is the thing leadership keeps asking for and rarely has.

The companies that get value from agents over the next few years will not be the ones that deployed the most of them. They will be the ones that knew exactly how many they had, what each one cost, and what each one returned, and were willing to turn off the rest. Sprawl is the default. A detox, run on a schedule, is the discipline that keeps the bill and the risk in line with the value.