When to stop adding AI agents and start consolidating the ones you run

This week the market shipped ways to manage agents, not more agents

I went through this week’s agent launches expecting the usual parade of new capabilities. That is not what showed up. Vorlon shipped a protocol-layer gateway that blocks or masks an agent’s action before a transaction completes. Couchbase shipped an Agent Catalog, a discoverable registry of the agents running against its data plane. Exabeam doubled its agent-behavior detections to 90 and open-sourced a telemetry library. Trust3 wired into Microsoft Copilot Studio to find shadow agents and shut them off in real time. Ory started baking agent identity into code at build time. All of it landed between July 1 and July 3, per the AI Agent Store roundup.

Read the list again. Not one of those is a new agent. Every one is a way to see, govern, or switch off agents that already exist. When the smartest tooling money in the market pivots in the same week from building agents to accounting for them, that is a signal about where the pain actually is.

Defense One reported on July 2 that GenAI.mil, the Pentagon’s generative AI platform, crossed 1.7 million users and more than 100,000 custom agents. One organization. Six figures of agents. Nobody sat down in 2025 and planned for a hundred thousand of anything. It accreted one useful little agent at a time, which is exactly how sprawl always arrives: never as a decision, always as a sum.

Why the default Series C move is to add one more agent

The instinct at this stage is understandable. Something works, so a team ships an agent for it. Another team sees that, and ships their own. Agentic tooling has made spinning one up a Tuesday-afternoon task, and this week’s OpenSage release from Berkeley’s RDI lab pushes that further still: agents that generate their own sub-agents and synthesize their own tools on the fly. Capability keeps getting cheaper to add.

So companies add. Forrester, in a piece last month, found that three-quarters of enterprise leaders say they are adopting agentic AI, while only a small minority have anything running in meaningful production beyond a chatbot with ambitions. The gap between “we have agents” and “we run agents” is where the trouble lives. Forrester named the reason well: a “trust tax,” where every autonomous action has to be logged and defensible to an auditor, and right now that cost is too high.

Here is what this week’s governance launches are really selling, underneath the feature lists. An inventory (Couchbase’s catalog). A control surface (Vorlon’s gateway). Eyes (Exabeam’s telemetry). A kill switch (Trust3). An identity for each agent (Ory). Strip the branding and it is the same shopping list a company writes when it has lost track of how many agents it is running and who owns them. The market built the fire extinguishers because it can smell the smoke.

AvePoint’s late-June State of AI report, run with Osterman Research across 750 enterprise leaders, put numbers on the smoke.

In that same study, 46.9% of employees now use agents weekly or daily, and 88.4% of organizations reported at least one agent-related security incident in the past year. The blind spot is widening faster than the fleet: the share of firms that cannot account for unsanctioned agent activity nearly tripled.

The binding constraint is accounting, not adding

Your agent count is a vanity metric. The number that matters is how many of them you can name an owner and a kill path for.

Here is the part that should change a Series C board conversation. Adding agents is a solved problem. The org already knows how to do it, cheaply, at will. The unsolved problem is the one nobody put on a roadmap: keeping a current list of what exists, who owns each one, what it is allowed to touch, and how fast it can be stopped. Gartner has reported that only about 18% of organizations keep a current, complete inventory of the agents already running inside their walls. That means four out of five are scaling something they cannot fully see.

For a company approaching enterprise sales, audit, or an eventual exit, that invisible fleet is the risk that keeps a CFO awake. One misbehaving agent with a valid credential and outbound access does not care that the other 400 are fine. This is the Series C nightmare in a single sentence: an incident from an agent nobody remembered owning, quietly undoing two years of trust built with enterprise buyers. Gartner expects 40% of enterprises to demote or decommission autonomous agents by 2027, mostly after an incident makes the governance gap visible. Retirement is coming to these fleets whether it is planned or forced. Planned is cheaper.

The counter-tension is real and worth naming out loud. The same week the control tools shipped, OpenSage shipped tooling to spawn more agents automatically. The sprawl vector is accelerating while the accounting vector races to catch up. Betting that the accounting will sort itself out later is the expensive assumption.

Consolidation is the scaling move, not a retreat from it

What the honest operators are doing looks less like a purchase and more like a quarter of cleanup. Build the register first: one list, every agent, a named human owner, the data and actions each is scoped to, and a tested way to switch it off. Then rationalize. Two agents doing the same job get merged. An agent nobody owns gets retired. An agent touching regulated data without a reason gets pulled. Forrester’s own advice was to scale in stages, behind approval gates and rollback paths, widening autonomy only when the controls earn it. That is not caution for its own sake. It is how a fleet stays survivable while it grows.

None of this requires the exotic. It is the same move good operators have always made when a system got away from them: stop, count what exists, name who owns each piece, cut what is not earning its place. AI did not change that discipline. It just made the pile grow faster.

The number I’d put on the Series C board slide

If I had one slide at the next board meeting, it would not carry the agent count. That number only ever goes up, and it flatters everyone in the room. I would put a harder number up instead: of the agents we run, how many have a named owner and a tested kill path. If the answer is 30% of the fleet, the honest headline is that we are running most of our AI on trust and hope. Getting it to 90% is not a technology project. It is a decision to treat the fleet like something that belongs to the company rather than something that happened to it. The teams that get through the next year cleanly will not be the ones with the most agents. They will be the ones who could still tell you, without flinching, what all of theirs were actually doing.