A recent McKinsey paper on agentic economics makes a point that slipped past most of the commentary about token prices: no autonomous system should run without a mandate, a budget, and a stopping rule.
Mandates and budgets we know how to write. Stopping rules are the interesting part, because most of us have never had to write one.
Traditional automation stops because the workflow says so. Record updated, invoice reconciled, report produced. There is an end state, and everybody recognises it when it arrives.
Now ask an agent for the strongest possible launch strategy, a proper read on a new market, or a creative proposition with some life in it. When is that finished?
It can always read one more source, test one more scenario, generate one more idea, or send a second agent to pull the first one’s work apart. Nothing inside the objective says when enough has been done.
“Done” was never really a property of the work. It was a property of the workflow, and the workflow is precisely what we have just removed. What survives is marginal value: stop when the next unit of effort stops changing the answer, rather than when the ground has supposedly been covered.
Focus On: the average is hiding the problem
A study of eight frontier models on coding tasks found that runs on the same task varied by as much as 30 times in token consumption. More consumption did not reliably buy accuracy. Performance often peaked at some middling cost and then flattened out. The models were also poor at predicting their own consumption before starting, and human estimates of task difficulty correlated only weakly with what the agents eventually spent.
The evidence comes from software engineering, and I would not lift those ratios straight into marketing. The mechanism travels well enough.
Two agents can take the same brief and return answers of similar quality by completely different routes. One reads three sources and stops. The other reads ten, calls two tools, hits a dead end, retries, and then spends four rounds marking its own homework.
That breaks the relationship between a unit of work and its expected cost, which is the relationship most planning assumptions quietly depend on. An average tells you very little about the tail, and the tail is where an autonomous system consumes intelligence without producing anything for it.
Research into multi-agent software development offers another clue. In one experiment, iterative code review accounted for 59.4 per cent of total token consumption. The first answer was cheap. Checking it was not.
Anyone who has watched a good deck get worse somewhere between version six and version eleven will recognise the shape of that.
Where “done” stops meaning anything
Updating a CRM record, classifying an enquiry, enriching a contact, or executing a structured campaign all have an observable finish. Research, competitive analysis, strategic planning, creative development, and scenario work do not.
At that end of the spectrum, autonomy starts to look like recursive consumption wearing the costume of improvement. The agent carries on because carrying on is available, not because the next pass is likely to matter.
A hard spending cap is the obvious remedy and, more often than people expect, the wrong one.
Anthropic reported that its multi-agent research system consumed roughly 15 times the tokens of a chat interaction and beat a single-agent system by 90.2 per cent on its internal research evaluation. On a separate browsing benchmark, token usage alone explained 80 per cent of the performance variance.
Sometimes the extra thinking is worth paying for. A fifty million pound acquisition analysis should be allowed to consume rather more intelligence than a five hundred pound administrative task. This is capital allocation, and it belongs closer to how you think about funding a pilot than to how you think about a software licence.
From a cap to a gate
Marginal value is an elegant idea until somebody has to encode it. Nobody is going to compute the expected return of the next search.
So you need observable proxies, and the practical shape is a gate.
Convergence between successive outputs suggests that exploration has plateaued. It is cheap and easy to automate, and it is limited: it measures agreement, not correctness. An agent can converge with real confidence on a poor answer, particularly when every iteration draws on the same sources and the same framing. Convergence tells you when to stop generating. It says nothing at all about when to stop checking.
Convergence should therefore trigger something rather than conclude something. What it should trigger is an independent critic with an adversarial brief: find the flaw, the missing evidence, or the alternative reading that would change the conclusion.
“Independent” has to mean something here. A critic running the same model over the same corpus with the same context will reproduce the same blind spots, and do it politely. Changing the prompt creates a second role, not a second perspective. Real independence usually costs something: a different model, different sources, different tooling, or a person.
If the critic finds nothing material, human review should apply above a defined threshold of value, risk, or irreversibility, and not below it. Put a person in front of every low-stakes output and you have paid for autonomy without buying it.
Confidence scores belong in the telemetry rather than in the decision. They can route uncertain cases, and over time they will tell you whether stated confidence bears any relationship to actual performance. Familiarity makes humans and machines feel more certain without making either more accurate. Log confidence. Do not take instructions from it.
Five lines in every mandate
This belongs in the operating mandate of each agent, not in a generic enterprise AI policy. A research agent, a creative agent, and a campaign-execution agent have different economics and deserve different boundaries.
Five lines:
Value at stake. What is a good outcome worth here, and how much failure can we absorb?
Cost envelope. What may this agent consume: models, tools, external data, other agents, and human attention, which is usually the expensive one.
Stopping rule. What evidence tells us that further exploration is now producing diminishing returns?
Verification rule. What gets checked, by whom or what, and how independent does that check need to be?
Escalation owner. When the agent stops without an acceptable verification, who accepts the result, intervenes, kills the work, or authorises more intelligence?
The fifth is the one that gets left blank. Without a named owner, stopping rules are decoration. The agent either fails quietly or keeps spending, because nobody has taken responsibility for ending it.
The question at the review meeting is no longer “what did this agent cost?” It is “what did the last unit of effort change, how do we know, and who can authorise another one?”
We are teaching machines to work from goals instead of instructions. Goals, unlike workflows, almost never carry their own termination conditions.
Take the one agent you already have in production and write its five lines before the quarter closes. If nobody will put their name against the fifth, what you have running is not an agent. It is an open tab.
Follow me
That’s all for this week. To keep up with the latest in generative AI and its relevance to your digital transformation programs, follow me on LinkedIn or subscribe to this newsletter.
Disclaimer: The views and opinions expressed in Chronicles of Change and on my social media accounts are my own and do not necessarily reflect the official policy or position of S&P Global. This issue cites published engineering research from Anthropic, a vendor with a commercial relationship with S&P Global.