The essay made an argument in plain language. This is the same argument with three things added:
- the mechanism named,
- the terms fixed to the stack they come from, and
- the evidence set out.
The loop, mechanically
An agent loop is a cycle with an explicit exit:
- The model is shown the goal and the work so far.
- It decides a single next action.
- That action runs, and its result is fed back.
- The cycle repeats.
The engineering art is entirely in making it stop on purpose.
In the Claude API the turn boundary is carried by a stop reason, and reading it correctly is what separates a working loop from a broken one:
| Stop reason | What the loop must do |
|---|---|
tool_use | The model is asking to call a tool. Execute it, return the result, continue. |
end_turn | The model considers the reply finished. The loop may exit. |
max_tokens | The output ceiling was reached. The reply is truncated: do not treat it as complete. |
stop_sequence | A configured sequence was emitted. |
Two defects come up again and again:
- Treating a
tool_usestop as a final answer. The tool call is silently dropped and a half-finished reply ships. - Ending on prose. The loop stops because the text contains "let me know if you need anything else." A loop that ends because the model sounds finished will one day not end at all, or will end early on a close that is not closed.
The working definition worth adopting, because it names every part that matters:
An agent is a model-directed execution loop that selects and sequences actions, uses tools to affect or inspect an environment, observes results, and adapts within bounded authority and explicit completion or escalation conditions.
And the line to keep for meetings: workflows decide the path; agents decide part of the path; tools do the work; policies decide what is allowed.
Is it even an agent? The control path test
"Agent" has been applied to almost everything, which makes the word useless for procurement. The test that survives contact with a vendor is the control path: at runtime, what decides the next step, coded logic or the model?
| Rung | Shape | Control path | Agent? |
|---|---|---|---|
| 1 | Deterministic automation | Fully predefined | No |
| 2 | Single model call | One request, one response | No |
| 3 | Retrieval assistant / RAG | Predefined retrieve-and-answer | No |
| 4 | LLM workflow | Predefined chain, router or parallel path | No |
| 5 | Bounded single agent | Model selects some steps and tools, inside limits | Yes |
| 6 | Multi-agent system | Several goal-directed agents coordinate | Yes |
Two things follow.
- The governing principle is minimum sufficient agency. The correct answer is the lowest rung that does the job, and rungs 1 to 4 are not consolation prizes.
- Most of what is sold as agentic sits at rung 3 or 4, where the control path is coded. That is not a criticism of those products. It is a statement about what you are budgeting for, because a fixed path and an open loop do not cost money in the same way at all.
The reconciliation in the essay is rung 5. Nothing decided in advance that the third item would fail to tie, and nothing decided in advance that the response would be an accrual proposal rather than a match.
The economics, in detail
The claim is that unit price and total spend are moving in opposite directions, and that the loop is the reason.
The mechanism. Recurring cost per unit of work is roughly:
- the number of turns,
- times the tokens per turn,
- times the price per token,
- plus the non-model cost each turn drags along: retrieval, tool round trips, and reviewer time on anything escalated.
Prompting fixes the first factor at one. A loop makes it a variable the model itself sets at runtime. A 95% fall in the third factor is swamped by a tenfold change in the first, and loops routinely take ten or more turns on real work.
The evidence, dated 2026-08-18.
| Finding | Source |
|---|---|
| Token costs fall ~95% by 2030 while agentic inference costs rise ~5× over two years: Gartner's "inference paradox" | CIO |
| Only 35% of C-suite leaders say AI consistently delivers outcomes, earns regulator confidence and allows control (HFS Research + TCS, 100+ executives) | CIO Dive |
| Only 11% of technology CxOs report being prepared for agent deployment at scale; two-thirds are accountable for AI systems they do not fully control | IBM |
| Anthropic annualized revenue reached $65B at the end of July, from $9B eight months earlier | TechCrunch |
That last row is the one people misread. A vendor's revenue is not evidence about your unit costs. It is evidence that total enterprise consumption is rising steeply while unit prices fall, which is the same paradox seen from the seller's side of the invoice.
Our own analysis of the corpus reaches the same place from two independent sets of evidence, both at high confidence:
- Total AI spend rises even as inference costs fall, because adoption scales faster than efficiency gains.
- AI bills are tripling while per-token costs decline.
A third finding, at medium confidence, notes that wasted tokens in agentic workflows and rising infrastructure costs together contradict the common story that agentic AI delivers near-term margin improvement.
The right unit is cost per accepted outcome, not cost per token, cost per call or cost per seat. A loop that takes twenty turns and produces a reconciliation a controller signs is cheaper than one that takes four and produces something reworked by hand. Token accounting flatters the second.
Why observability is a finance control
"An unobservable loop is an unbudgetable one" is not rhetoric.
Every quantity above is per run: turns taken, tokens per turn, tools called, escalations raised, outcomes accepted. None of them can be read off an invoice, because the invoice adds up every loop you run. If a trace does not record them per run, the only cost signal left is a single monthly number that cannot be broken down. That is exactly the state the field reports: enterprises cannot compute a return because they do not have a complete picture of the cost.
A trace good enough to budget from records, at minimum:
- the run identifier,
- the skill loaded,
- each tool call with its outcome,
- the turn count and the stop reason,
- tokens in and out,
- any escalation, and
- whether the result was accepted downstream.
That is not an exotic ask. It is the agent equivalent of a general ledger entry, and it is the difference between an AI line item and an AI cost model.
The six parts, in the stack's own words
| Component | What it is | Where the term comes from |
|---|---|---|
| The loop | Model-directed cycle with an explicit stop condition, driven by stop reasons | Agentic architecture |
| Tools | Described, bounded interfaces to systems and data; MCP is the open standard for exposing them | Tool design; MCP is the agent-to-tool layer |
| Skills | Packaged procedure in its own file: a name and a "when to use me" description over a body loaded on match | Progressive disclosure |
| Governance | Policy, permission, approval and prohibition, enforced outside the model | Bounded authority |
| Memory and state | Durable run state keyed to a run id, so an interrupted run resumes | Resumability |
| Observability and evals | Traces of what happened, and tests of whether it was right | Evaluation and reliability |
One boundary is worth keeping straight, because vendor material blurs it all the time: MCP is the agent-to-tool layer. Agent-to-agent coordination is a different layer with different protocols. Mix them up and you can no longer reason about which risks and which costs sit where.
What this series will not claim
Our own reading of the evidence is that AI productivity is currently going to a small group of builders and early adopters rather than to the typical enterprise employee. That makes broad workforce productivity figures misleading for at least the next twelve months. Nothing here should be read as a claim that a finance function can stand up a portfolio of loops in a quarter.
The claim is narrower and, we think, more useful: the loop is the unit that can be designed, controlled, observed and priced. On the evidence, the organizations building control into their loops now are already ahead of the ones waiting for a governance framework to mature.
⚠️ Every figure in this analysis is a 2026-08-18 reading, and the Gartner projections run to 2030. When one moves, this page records the change and the date it changed.
The framing of this series is inspired by dadloop, an agent harness built and published by Swami Chandrasekaran, Partner and Global Head of AI & Data Labs at KPMG (LinkedIn · github.com/swamichandra/dadloop). The six-part anatomy, the traced run and the named-skill catalog are his teaching devices; the finance domain, the loop library and every figure here are ours.