Anyone who lived through the enterprise move to cloud already knows how this story goes.
In the early years, cloud was magic. You could spin up infrastructure in minutes instead of months. No more capacity planning, waiting on a data center, capex committees. You needed a server, you got a server, or ten, or a hundred, on demand, and you only paid for what you used.
Then the bills started arriving.
The very thing that made cloud easy to adopt (frictionless, on-demand, pay-as-you-go) turned out to be the thing that made it easy to overspend on. A fixed, planned line item became variable and sprawling. Forgotten instances ran for months, environments were oversized “to be safe”, dev sandboxes never got turned off, and egress charges nobody had modeled showed up out of nowhere. When finance finally asked the obvious question, who is spending all of this, and on what, the honest answer, for an embarrassingly long time, was: we're not sure.
That question is what eventually created the entire discipline of cloud FinOps. Tagging conventions, cost-allocation reports, showback and chargeback, rightsizing, reserved capacity, whole teams whose job was to reconnect spend to the people and applications that caused it. The conversation matured from “is cloud worth it?” to “are we running cloud responsibly?”
We are about to run that exact play again with agents. Only faster.
Agent costs: Same arc, shorter fuse
§ 01 — The cloud cost curve, replayed at agent speed
Right now, agentic AI is in its magic phase. Agents do real work, EVERYONE is piloting something, and some agents are already making their way into production workflows. You can already see the next phase coming.
The same properties that make agents easy to adopt make them easy to overspend on. There is a token meter, and usage is rising exponentially — often encouraged by leadership. Part of what's driving this usage increase is that a single human request doesn't map to a single action anymore; it fans out into a tree. A planner agent spawns sub-agents, each sub-agent reasons and calls tools, and each tool call can trigger more reasoning downstream. Tokens are consumed at every node, on demand, with no natural friction telling anyone to stop. The spend is variable, it sprawls, and it compounds in unexpected ways.
A single human request on the left fans out to the right into a planner agent, then into two sub-agents, then into four tool calls. Each agent, sub-agent, and tool call carries a token cost, and a running total climbs as the tree expands — illustrating that one request burns tokens at every node, so the spend is variable, sprawling, and compounding.
When the first serious agentic bill lands, the question will be identical to the cloud one: who is spending all of this, and on what? And if nothing changes, the answer will be just as uncomfortable: we're not sure.
The difference is the clock. It took the industry the better part of a decade and a great deal of wasted money to build cloud cost discipline. Agents are scaling faster than cloud ever did. The gap between “this is amazing” and “what did we just spend?” “why did we spend so much and on what did we get for it?” is going to be measured in quarters, not years.
Retrofitted attribution: The mistake we don't have to repeat
§ 02 — Why cloud attribution was always bolted on after the fact
Here's the part worth slowing down for, because it's where the analogy stops being a warning and starts being an opportunity.
Cloud cost attribution was always retrofitted. The billing layer knew that an instance ran for 400 hours; it did not know which team, which application, or which business purpose that instance served. So, enterprises bolted attribution on after the fact: tagging conventions everyone was supposed to follow, third-party tools reconstructing intent from billing exports, analysts working through historical spend. It partially worked, but the source of truth for spend and the source of truth for intent were two different systems, and the intent one was not wired into the spend one at the moment the spend happened.
Cloud cost discipline was built backward, reconstructed from billing exports years after the money was gone, because the systems that spent the money and the systems that knew why were never the same systems. Agents do not have to inherit that. The arc is the same and the fuse is shorter, but for once we can see the bill coming. We have the opportunity to build attribution in from the start, at the one layer that already knows who acted, on whose behalf, and why.
That layer is the control plane. The next post is about a key element of data that lives in the control plane: the delegation chain, and why it turns out to be the cost chain too.
Explore Related SecureAuth Solutions
See the agentic bill before it lands.
SecureAuth Agent Authority is the enterprise control layer for autonomous AI — identity, scoped delegation, approval gates, audit, and runtime enforcement in one platform. Every agent. Every action.
About SecureAuth
SecureAuth provides identity and access management solutions that enable enterprises to implement customized, resilient authentication infrastructure. Through Continuous Authority, flexible deployment options, and deep composable capabilities, SecureAuth helps organizations defend against modern identity threats while maintaining usability and operational efficiency.