Everything covered so far in this series assumes a cost model that is at least linear: more usage costs proportionally more, and the relationship, while unmanaged, is at least predictable. Agentic AI breaks that assumption. An agent that plans, calls tools, evaluates its own output, and re-plans when the result falls short does not spend tokens once per user request. It spends tokens across every step of a reasoning loop that can extend for many iterations, and a second agent delegating work to a third multiplies that loop again. Enterprises scaling agentic AI without the observability, optimisation, and governance disciplines from earlier in this series are scaling a cost problem that compounds faster than any of them can retrofit.
Why Agentic Cost Does Not Scale Like a Single Call
A conventional AI feature has a cost profile that is straightforward to reason about: one request, one model call, one token cost, repeated at whatever rate users invoke it. An agentic workflow has none of that predictability. A single user request can trigger a planning step, several tool calls, an evaluation of whether the tool output actually answered the question, a re-planning step if it did not, and a final synthesis step, each of which is its own model call with its own token cost. Add a second agent that the first agent delegates a subtask to, and the same pattern repeats inside that delegation, nested inside the cost of the outer loop.
This is not a defect in agentic architecture. It is what makes agents capable of handling tasks a single-call system cannot: the ability to check its own work and adapt when the first approach does not succeed. The capability and the cost multiplication are the same mechanism, which means the honest framing is not “fix the multiplication” but “meter and bound it deliberately,” because the alternative to token multiplication in agentic systems is frequently the loss of the capability that makes them worth building in the first place.
Where the Multiplication Becomes a Liability
The gap between capable and unbounded is where enterprises get into trouble. An agent with no limit on re-planning iterations, given a task it cannot cleanly solve, will keep trying, and every attempt has a token cost. An agent architecture with no visibility into how many tool calls a typical task requires cannot tell the difference between a workflow that usually completes in three steps and one that occasionally spirals into thirty, until the invoice reflects it. A multi-agent system with no per-agent cost attribution cannot tell which agent in the chain is driving the majority of the spend, which makes it impossible to target optimisation effort at the part of the system that would actually move the number.
None of these are new problems in kind. They are the observability and optimisation disciplines from earlier in this series, applied to a workload where the absence of that discipline is punished faster and more severely than it is in a conventional AI feature, because the cost multiplication compounds within a single user interaction rather than accruing gradually across many separate ones.
The Prerequisites, Not Afterthoughts
This is why the sequencing in this series has not been arbitrary. Real-time token observability, applied to an agentic system, means attributing cost not just to the feature but to the specific agent and the specific step within a multi-step workflow, so that a cost spike is traceable to the part of the system actually causing it. Optimisation discipline, applied to agentic architecture, means the same model right-sizing and context engineering from earlier, plus explicit bounds: maximum re-planning iterations, maximum delegation depth, defined escalation to a human or a cheaper fallback path when an agent is not converging on an answer efficiently. Governance and ownership, applied here, means the team deploying an agentic workflow is accountable for its cost behaviour under realistic and adversarial conditions, not just its behaviour in the demo that got it approved.
An enterprise that tries to build agentic AI capability without these three disciplines already in place is not choosing to skip a step. It is choosing to discover the cost of skipping it after the workflow is already in production and difficult to unwind, at exactly the scale where unwinding it is most expensive.
The Capability Enterprises Actually Want
None of this is an argument against agentic AI. It is an argument that the enterprises who get the most value out of agentic capability will be the ones who treated token economics as infrastructure to build before scaling agentic workloads, not as a problem to solve after the fact. The same architectural discipline that makes an agent trustworthy enough to act with real autonomy, bounded reasoning, clear escalation paths, verifiable outputs, is the discipline that keeps its cost bounded too. The organisations that build both together will be the ones running agentic AI at scale with confidence. The ones that built agentic capability first and are only now asking what it costs are the ones for whom this series will read less like a preview and more like a diagnosis.
