Getting Better at Getting Better: What Recursive Self-Improvement Would Mean for Martech
AI labs are building agents that improve the way they improve. What would that look like in marketing, where the test is not a benchmark that runs in seconds but a customer who takes weeks to respond?
The argument in one line. Recursive self-improvement in martech means a system that improves not only its marketing decisions but also the agents that make them, and eventually the way those agents are improved. Every competitor will have access to the same models, so the ability to propose improvements will be plentiful. The ability to verify them will be scarce: a fast and honest evaluation clock, an objective the system cannot game, and the rights and evidence to test inside real brands. The moat is improvement velocity: the capacity to generate trustworthy evidence faster and more cheaply than competitors, and to turn it into better agents. It rests on verification, not on intelligence.
1 Getting Better at Getting Better
In May 2025, Sakana AI and researchers at the University of British Columbia described a coding agent that rewrites its own code. The Darwin Gödel Machine proposes changes to itself, tests each variant on programming benchmarks, keeps the versions that score better and builds on them. Over the course of its run, its success rate on a widely used benchmark of real software bugs rose from 20% to 50%. The same month, Google DeepMind unveiled AlphaEvolve, which pairs language models with automated evaluators to discover better algorithms. A heuristic it found for scheduling work across Google’s data centres had by then been running in production for more than a year, and recovers about 0.7% of the company’s worldwide computing capacity. The technique has already reached marketing. In May 2026, DeepMind reported that WPP had used AlphaEvolve to refine the models it uses to predict campaign performance, achieving 10% accuracy gains over its own manual optimisation.
In March 2026, researchers at Meta went a step further. Their Hyperagents make the improvement process itself editable, so that the system improves not only how it performs a task but how it searches for improvements. This is what the field calls recursive self-improvement: a system that gets better at getting better.
These systems share an advantage that is easy to overlook. Their test is fast. Code runs or fails, a benchmark scores in minutes, and the loop can turn thousands of times. The Hyperagents authors make a further, subtler point. The Darwin Gödel Machine worked partly because evaluation and self-modification were the same kind of task: becoming better at coding made the agent better at rewriting its own code. Outside coding, that alignment breaks. Becoming better at reviewing research papers does not make an agent better at editing the software that does the reviewing.
Marketing is an extreme case of that break. A marketing agent that changes how it approaches dormant customers may wait sixty days to learn whether the change worked, and the answer comes from customers, not from a benchmark. The question for martech is not whether self-improving agents are coming. They are. The question is what self-improvement looks like when the test is slow, noisy and human.
This matters now because agents have become the centre of every martech roadmap. Nearly every vendor now offers agents that build journeys, write content or recommend the next action, and the features are converging. In Martech: The Moat Is the Loop, I argued that the defensible advantage lies not in any feature but in a loop that compounds with use. Self-improving agents raise the stakes of that argument. If the agents themselves can be improved by machines, and the machines that improve them can be licensed by anyone, the question becomes which company’s improvements can be trusted fastest, and why. Note what WPP’s gain measured: the accuracy of a prediction model on campaign data, a fast and automatic score. That is a real advance. It is not yet evidence that customers bought more because of it, and the gap between those two things is the subject of this essay.
My answer, which this essay develops, is that self-improving marketing is possible, but it will be bounded, governed and paced by verification. It will not be an explosion. It will be a steady acceleration, and the companies that achieve it will be the ones that can verify improvements fastest without fooling themselves.
In software, the test runs in seconds. In marketing, the test is the customer, and the customer takes weeks.
2 Four Levels of Improvement
“AI marketing” now covers so many things that it helps to separate four levels of improvement.
The first is automation. Agents do existing work faster: they assemble campaigns, write variants, build segments and schedule sends. The decisions are no better than before; they simply take less time and fewer people.
The second is optimisation. The system makes better decisions as evidence accumulates: which customers to contact, which offer to make, when to send. This is the familiar promise of machine learning in marketing, and when it is done properly, with holdouts, it is valuable.
The third is self-improvement. The system changes its own machinery: the decision rules, prompts, tools and workflows of its agents. It proposes a new version of an agent, tests it against the current one, and keeps whichever wins. The object of improvement is no longer the campaign but the thing that makes campaigns.
The fourth is recursive improvement. The system improves the process by which it discovers, builds and evaluates changes to its agents. It learns which kinds of change tend to work, which tests are worth running and how to propose the next change better than the last.
An example makes the ladder concrete. A system that finds a better subject line is optimising. A system that finds a better way of choosing subject lines, and changes its own choosing process, is improving itself. A system that finds a better method for generating and testing new choosing methods is improving its ability to improve. The recursion lives in that last step.
| Level | What improves | Example | How it is verified |
| 1. Automation | Speed of execution | An agent assembles and schedules a campaign | Time and cost saved |
| 2. Optimisation | Marketing decisions | Choosing which offer each customer receives | Lift against a holdout |
| 3. Self-improvement | The agents themselves | Rewriting the rules for choosing offers, and testing the new version | Champion against challenger on live, randomised groups |
| 4. Recursive improvement | The process of improvement | Learning which kinds of agent change tend to work, and proposing better ones | Verified gains arriving faster across successive cycles |
Figure 1. Four levels of improvement. Most of what martech sells as AI today sits at the first two.
Most of what martech sells as AI today sits at the first two levels, and a good deal of it at the first. That is not a criticism: automation and optimisation are useful, and the higher levels are impossible without them. But they do not compound in the way the higher levels could.
A system that finds a better subject line is optimising. One that finds a better way to find subject lines is improving itself.
3 Three Nested Loops
In The Moat Is the Loop, I described two loops that could give a martech company a durable advantage. The evidence loop makes each deployment more likely to work, as verified results across brands improve the next prediction. The replication loop makes each deployment cheaper, as successful field work becomes reusable product. Both are about improving the marketing and the way it is installed. Self-improvement adds structure around them.
Picture three loops, nested. The innermost is the marketing loop, at the level of the customer: decide, act, measure against a holdout, learn. It improves the intervention. Around it sits the agent loop, at the level of the system: diagnose where an agent falls short, redesign it, test the new version and promote it if it wins. It improves the machinery that makes decisions. The outermost is the improvement loop: study which agent changes worked and which failed, and improve the way future changes are proposed and tested. It improves the process of improvement.
Each outer loop feeds the one inside it. Better methods of improvement produce better agents sooner; better agents produce better marketing decisions; and every marketing decision, properly measured, produces evidence for all three loops at once.

Figure 2. Three nested loops. Each outer loop improves the one inside it, and every measured decision feeds all three.
A hypothetical example shows how this could unfold over five generations of a recovery programme for a retailer.
In the first generation, an agent identifies customers who have been dormant for between 90 and 180 days and makes them offers based on their purchase history. Holdouts show which offers create incremental contribution.
In the second, the agent learns that its discounts are being spent on customers who would have returned anyway. It redesigns its policy: useful content for one group, a reminder to restock for another, and discounts only where the extra incentive pays for itself.
In the third, reviewing several experiments, the system notices that most of the improvement came not from better messages but from choosing different customers, including deciding whom to leave alone. It changes the agent that selects whom to contact.
In the fourth, the system observes that experiments on customer selection have produced more reliable gains than experiments on content, and shifts its testing effort accordingly.
In the fifth, working across many brands, the improvement loop learns which kinds of agent change carry over from one brand to another and which need local calibration. It begins to propose better starting agents for each new brand, and to test them more efficiently.
By the fifth generation something qualitatively different has happened. The system is no longer applying lessons from earlier campaigns. It is applying lessons about how to build marketing systems. That is the recursive step. It is also, as section 7 will argue, where the first danger appears, because the fourth generation is exactly how a self-improving system can talk itself out of its best ideas.
The first loop improves the marketing. The second improves the agents. The third improves the way the agents are improved.
4 The Evaluation Clock
Every self-improving system is limited by its evaluator. AlphaEvolve’s evaluators score a candidate algorithm automatically. The Darwin Gödel Machine runs each new version of itself against a suite of programming tasks. In both cases the verdict is cheap, quick and unambiguous, which is why the loops can turn so fast.
In marketing, the evaluator is customer behaviour, and customer behaviour is slow, noisy and costly to observe. A system can propose a thousand changes to its agents in an afternoon. It cannot verify them faster than customers respond, and it cannot run more experiments than it has customers and attention to spend on them.
It helps to think of the speed of improvement as the product of two things: how quickly the system can learn whether a change worked, and what share of its proposed changes turn out to be improvements. The second depends on the quality of the proposals, which is where better models help. The first depends on the evaluation clock, and in marketing the clock dominates.
The clock does not limit how many experiments a company can run. It can run a hundred in parallel, on different cohorts, with their results arriving in overlapping windows. What the clock limits is learning that builds on learning. An improvement that depends on the result of an earlier experiment must wait for that result, and the next improvement must wait for that one. Recursive improvement is exactly such a chain: each better agent is designed using what the previous one revealed. It is the length of each link in this chain of dependent learning, not the number of experiments launched, that sets the pace.
The arithmetic follows. If each link takes sixty days to read, the chain advances about six links a year. If each link takes two weeks, it advances twenty-six. Over three years that is eighteen links against seventy-eight. Running more experiments in parallel widens each link, testing more ideas at every step, but it does not shorten the chain. And parallelism has costs of its own: more simultaneous tests mean smaller groups and noisier results, experiments can interfere with one another, and a customer contacted by several of them is not a fresh subject for each.
There is also more than one clock. A completed form can be read at once; a repeat order takes a week; a renewal may come round once a year. And a retention programme can look excellent after thirty days and turn negative after six months, because it has trained customers to wait for discounts. A self-improving system needs to know which clock belongs to which promise, and must not promote a change on the strength of a clock that has not yet finished ticking.
So the strategic question is how to speed up the clock without making it dishonest. There are four ways, and each comes with a condition.
The first is dense daily signals. A brand that sends one promotion a week to an inattentive list receives one weak signal a week, most of it silence. A daily email that gives the reader something to do, whether a quiz, a poll, a prediction or a card to collect, produces engagement signals every day from every reader who opens it. I have argued in earlier essays that such emails rebuild the attention brands have lost. Here they play a second role: they turn the inbox into a laboratory with a fast clock. Changes to selection, timing and content can be read in days rather than months. I have written about earning that repeated attention in The Next Open. The condition is twofold. What matters is useful evidence density, not event volume: automated opens by privacy features and repetitive taps can be plentiful and tell the system almost nothing. And engagement is not contribution, so these readings must be checked regularly against holdouts that measure money.
The second is validated leading indicators: early measures, such as engagement in the first week, a return visit or a completed interaction, that have been shown against holdouts to predict contribution ninety days later. The system can iterate quickly on the leading indicator while the slow truth checks, at intervals, whether the indicator still predicts. The condition is that when an indicator stops predicting, the system must slow down until a new one is validated.
The third is calibrated simulation. Synthetic customers built from language models can screen ideas before any real customer sees them, and the evidence on how far to trust them is now accumulating. Amazon researchers tested simulated A/B tests against 67 historical marketing experiments. Uncalibrated, the simulations systematically overstated the size of effects; calibrating them against data from before each test cut the magnitude error sharply. The simulations matched the direction of the real result in about 70% of tests, but the authors caution that, judged against simple uninformed baselines on such noisy experiments, this is not clear evidence that the simulations had picked up a real directional signal. A separate team of statisticians, working with news-headline experiments rather than commercial campaigns, found that uncalibrated model responses recovered only 39% of the effect measured on people, and drew a sharp distinction: tests on simulated customers are correct only by assumption, while tests on people are correct by design. Simulation can be calibrated in promising ways, but its usefulness for choosing between interventions is not yet established. The condition is that it serves as a filter for rejecting weak ideas cheaply, never as a substitute for verification.
The fourth is replaying past decisions, and it needs the most careful handling. A record of decisions and outcomes allows a new agent to be run against past customer situations before it goes live. That is excellent for catching errors: a new agent that would have breached a permission, overspent a budget, contradicted a rule or behaved erratically can be rejected without touching a customer. AlphaEvolve’s scheduling heuristic was itself first tested in a simulator built from historical workloads. But replay is much weaker at estimating profit. If the old agent offered a customer a discount, we know how that customer responded to the discount; we do not know how they would have responded to something else. Replay can estimate the value of a new policy only under particular conditions: the old system must have assigned its actions with known, logged probabilities, it must have tried the new policy’s actions often enough on similar customers, and the statistics must be handled with care. Where the old system never tried an action on a group of customers, the history cannot say what that action would have done. The condition is that replay is used to reject invalid policies and test behaviour, and to estimate outcomes only where past experiments cover the ground.
| Method | Speed | Good for | Limits |
| Randomised holdout | Days to months | Establishing what was caused, in contribution | Slow, and costs attention; but the final authority |
| Daily engagement signals | Days | Quick reads on selection, timing and content | Engagement is not money; volume is not evidence; checked against holdouts |
| Validated leading indicators | Days to weeks | Fast iteration between holdout reads | Can stop predicting; re-validated regularly |
| Calibrated simulation | Minutes to hours | Screening out weak ideas cheaply | Overstates effects uncalibrated; direction not yet clearly better than simple baselines |
| Replay of past decisions | Hours | Catching errors and rule breaches; estimating value where past tests randomised the same actions | Cannot say what untried actions would have done |
Figure 3. Ways to speed up the evaluation clock. Fast methods propose and filter; the randomised holdout decides.
Each accelerator is a proxy, and proxies drift. The design principle that follows is simple to state and hard to keep: fast evaluators propose and filter; the randomised holdout decides. A self-improving system may use every shortcut available to discard bad ideas quickly, but nothing is promoted into the product on the strength of a shortcut alone. And the clock cannot be shortened indefinitely. A mature system will sometimes decide not to test another small variation, because the information it would yield is worth less than the customer attention it would consume. That is not a failure of ambition. It is discipline.
Self-improvement in marketing moves at the speed of honest verification.
5 The Evaluator Is the Moat
Recursion has a property that ordinary optimisation lacks: it amplifies the objective. A system that improves itself becomes steadily better at whatever it is rewarded for, and better at finding ways to be rewarded. If the reward is opens and clicks, it will learn to exhaust customers. If the reward is attributed revenue, it will learn to give discounts to people who would have bought anyway, the trap I illustrated in Martech: The Moat Is the Loop. Consider an agent rewarded on recovered customers alone. It will discover that larger discounts recover more customers, and each generation will offer a little more. Every report will show rising recoveries. The contribution, which nobody asked it to protect, will fall. A human team might drift into these habits slowly and occasionally. A self-improving system will pursue them systematically, quickly and with great ingenuity.
Recent marketing research offers a small but telling example. A team from Penn State, Rochester and Tennessee, publishing in the Journal of Marketing Research, built a framework that uses a company’s own campaign history to screen AI-generated email content. When they asked a language model to choose its own five best emails, the company’s data predicted those choices would perform poorly. Their answer was to divide the work: the model writes, and the company’s evidence judges. That division of labour is the principle on which self-improving marketing has to be built.
The right objective is verified incremental contribution: the extra profit an intervention causes, measured against a pre-agreed baseline and a randomised holdout, after the cost of incentives and fulfilment. Two kinds of protection sit alongside it, and they should not be confused. Some harms belong in the measurement: attention decay, the shrinking of the engaged base, unsubscribes and complaints. A change that lifts contribution this quarter while eroding the base of customers who still pay attention is not an improvement, and the evaluator should count the erosion against it. Other limits are not costs to be traded at all. Customer permissions, consent rules and contractual commitments are hard constraints: an action that breaches them is never taken, however much contribution it promises, and a change that would breach them is never promoted. A reward function that merely penalised such breaches would invite a self-improving system to find the ones worth paying for.
This gives outcome pricing a role larger than commerce. In No Lift, No Bill, I described a contract in which the vendor earns only on verified lift above an agreed baseline. In a self-improving system, that contract becomes the fixed point the loop is not allowed to change. The baseline, the holdout design, the formula for contribution and the measurement windows are set by people and written into agreements. The agents can propose changes to almost everything else.
That is the independence principle. The evaluator must sit outside what it evaluates. The research systems already work this way: both the Darwin Gödel Machine and Hyperagents were run with sandboxing and human oversight, and their scoring sat beyond the reach of the code being changed. In marketing the stakes are higher, because the system is acting on real customers and real money. An agent that could edit the rules that judge it would soon find that the easiest way to improve is to change the rules.
Why does this make the evaluator the moat? Not because of the scoring formula. Incremental contribution, holdouts and guardrails are not secrets, and any competitor can write them down. What is hard to copy is everything needed to run that formula well, at scale and at low cost: the rights to run experiments inside many brands, holdouts brands have agreed to, a permissioned history of decisions and outcomes over long periods, calibration to each brand’s circumstances, operations reliable enough to act on the results, and the funding to keep testing. Every competitor will have access to the same frontier models, and WPP’s experience shows that the tools for improving models are already commercially available. So the ability to propose changes will be widely shared. What will not be shared is the capacity to generate more trustworthy evidence, at lower cost, than anyone else. A rival with the same models and without that capacity can generate ideas as quickly. It cannot learn which ones work as quickly.
In a self-improving system, the evaluator is the moat: not the formula, but the capacity to produce evidence others cannot.
6 The Agent Factory
Software companies learned long ago that shipping code safely needs a pipeline: build, test, stage, release, monitor, roll back. A self-improving marketing system needs the equivalent for agents. Call it an agent factory.
Candidates enter the factory from diagnosis. An improvement agent studies the performance of the working agents and finds their weaknesses: a strategy agent that consistently overestimates the value of discounts, a creative agent that produces too many near-identical variants, an experimentation agent that spends test traffic on questions whose answers are already settled. It proposes changes to each. The candidates are tested first in a sandbox and against replays of past decisions. Survivors run in shadow mode, making decisions alongside the current agent that are recorded but not executed. Only then do they face the current version in a live, randomised champion-and-challenger test. Winners are promoted, versioned and monitored; losers are archived, not discarded. Every agent has a lineage, and any version can be rolled back.
Not every agent in the factory needs to be a language model. Most of the machinery should remain conventional computation: eligibility rules, uplift models, assignment systems, cost accounting, statistical tests and release controls. Language models earn their place where the system must interpret ambiguous evidence, propose alternative designs, write tools or reason about unfamiliar situations. They are not a substitute for a reliable experiment or an auditable record of what happened.
Proposals to change the improvement process itself travel through the same gates. An improvement agent may suggest searching a different class of problems, combining specialist agents differently or ranking ideas for testing in a new way. Each such change must earn promotion like any other. In particular, no agent can bypass the evaluator by proposing that a new evaluator would be smarter. The recursion operates inside a fenced factory, not outside it.
Promotion should be staged in the same way. A change that wins in one brand goes next to a handful of similar brands, and only then to everyone, with the guardrail measures watched at each step and an automatic stop if any of them moves the wrong way. The rate at which changes can be promoted should itself be limited, so that the system never changes faster than people can understand what it has done.
The archive matters more than it seems. The Darwin Gödel Machine keeps a library of its earlier agents, not only the best one, because a version that loses today may be the stepping stone to something better tomorrow. A marketing factory should do the same.
The factory also needs a clear boundary between what the agents may change and what only people may change.
| What agents may change | What only people may change |
| Decision policies and selection rules | Brand objectives and priorities |
| Prompts, content generation and tool use | Consent records and customer permissions |
| Workflow sequences and timing | Financial limits and incentive budgets |
| How test traffic is allocated | Contract terms, baselines and the vendor’s share |
| Proposals for new specialist agents | Measurement rules, holdout design and windows |
| Retiring weaker agent versions | The exploration budget and the evaluator itself |
Figure 4. The boundary. Agents may change how the work is done; only people may change what counts as success.
This changes the role of the people involved. In the earlier essay, Forward-Deployed Marketers built gravel roads inside each brand, and a product team paved them into highways for the next brand. In the factory, many of the paving proposals come from the machine. The Forward-Deployed Marketer’s job moves up a level: judging which proposals deserve trust, defining what better means for a particular brand, handling the exceptions the agents cannot, and keeping the brand’s judgement inside the loop. People set the objectives, the permissions, the economic limits and the standards of proof. The machine searches within them. The Software Foundry series described agents that write and maintain software; the agent factory applies the same idea to the agents themselves. And the replication test from Martech: The Moat Is the Loop applies with extra force: if the factory’s tenth improvement costs as much as its first, there is no factory.
The marketer stops designing every agent and starts deciding what counts as a better one.
7 The Exploration Trap
Return to the fourth generation of the recovery example. The system noticed that experiments on customer selection produced more reliable gains than experiments on content, and shifted its effort towards selection. Each step of that reasoning is sound. Its cumulative effect can be corrosive.
A self-improving system learns not only what works but what is easy to verify. Changes to selection read quickly and cleanly. Changes to content, new mechanics and unfamiliar ideas read slowly and noisily. Over time the system drifts towards what it can measure and away from what it cannot. Every individual decision is rational. The result is a system that converges on a narrow, safe and slowly declining set of tactics, and stops discovering anything new.
The same Journal of Marketing Research study shows why this matters. The framework flags messages that are too unlike past campaigns to predict reliably, and directs them to live testing instead. In its data, an unusual birthday greeting, which the framework could not predict, turned out to have the strongest effect on spending. A system that promoted only what it could predict would never have sent it.
The research systems point to the remedy. The Darwin Gödel Machine’s archive of diverse agents exists precisely to avoid settling too early on a good-enough answer. In marketing, the equivalent is a protected budget for exploration: a share of test traffic and experiment slots reserved for novel ideas, set by people and outside the system’s control, so that the system cannot quietly reallocate it to safer bets. Novel ideas should be given longer measurement windows, because their effects are slower to read. Failed and untested ideas should be kept as stepping stones. And people should keep feeding ideas in, from Forward-Deployed Marketers, from brand teams and from customers themselves, because the most valuable ideas are often the ones no model trained on the past would propose. The protection should be disciplined, not romantic. An untested idea has no right to reach a million customers. It has a right to a small, well-designed experiment with a bounded downside. And the system’s own scorecard must not punish boldness: if the factory is rewarded only for the share of its proposals that survive the holdout, it will learn to propose safe, incremental changes and achieve excellent statistics by refusing to ask difficult questions. Survival rates should always be read alongside the novelty of what was attempted.
A system that trusts only what it can predict will optimise away its best ideas.
8 The Financing Loop
Self-improvement costs money. Experiments consume test traffic, holdouts, computing, and the time of the people who judge the results. Who pays for the search?
Under the billing meters martech has used for twenty-five years, nobody has much reason to. A vendor paid per message sent or per user tracked earns the same whether its agents improve or not, and many improvements, such as better targeting or fewer wasted sends, reduce the volume it bills. I described those meters in Martech Beyond Messages and MAUs. They reward activity, and a self-improving system that made activity more efficient would shrink its owner’s revenue.
Outcome pricing reverses the incentive. When a vendor earns a share of verified lift, every improvement its agents discover raises its earnings. Verified lift earns Carry; Carry funds experimentation capacity; experimentation produces better agents; better agents produce more verified lift. Profit from past outcomes pays for the search for future ones, and the search becomes part of the business model rather than a cost centre. The brand gains too, because the improvements are found on its behalf and it keeps most of the lift they produce. The arrangement works only if that remains true: a vendor whose share of the gains rose faster than the brand’s would soon find brands unwilling to keep the experiments running.
There are limits. More experimentation is not automatically better. Every test consumes customer attention, and every holdout means some customers do not receive a treatment that might have helped them, which is a real cost to the brand. The financing loop therefore needs rules of its own: an experiment budget for each brand, a ceiling on the share of customers in tests at any time, priority for the experiments whose answers are worth most, and the capital discipline that any outcome business needs.
Profits from verified outcomes fund the search for the next ones.
9 Where to Start
If the evaluation clock governs the speed of self-improvement, the place to begin is where the clock is shortest.
That means workflows whose outcomes arrive in days: unfinished transactions, such as abandoned checkouts and incomplete applications; renewals due this month; onboarding and verification journeys; and the second purchase after a first. Results arrive quickly, individual decisions are small and reversible, and the stakes are bounded. Recovering deeply dormant customers is the larger prize, but its outcomes are sparse and slow. It is a place to arrive at, not to start.
A first pilot could run for ninety days on randomised, comparable groups of customers, with the same measurement rules and budgets for each. One group would be served by a fixed agent, maintained by people in the usual way. A second would be served by the same starting agent together with an improvement process allowed to propose, test and promote changes within the boundaries described in section 6, on a logged release schedule. A third would receive no intervention, as a control for both.

Figure 5. Two experiments. The first, a ninety-day pilot, tests self-improvement; the second, longer and across several cycles, tests recursion.
Before the pilot begins, the brand and the vendor need to agree a short list of things: the permissions the agents may use, the size and design of the holdouts, the boundary between what agents may and may not change, the guardrail measures and their thresholds, and an off-switch that the brand controls. None of this is exotic. It is the same discipline any outcome contract needs, applied to a system that changes itself.
The first question is whether the self-improving group delivers more incremental contribution per unit of operating cost than the fixed group, without more incidents, reversals or complaints. Analysts must resist presenting a sequence of selectively chosen winning tests as one clean result; the comparison is between the two groups as wholes, under an analysis plan fixed in advance. A yes shows that self-improvement adds value over a competent agent.
It does not show recursion, and the distinction matters because recursion is this essay’s distinctive claim. A single successful upgrade is not evidence of recursion, and nor is a faster improvement cycle, which might have come from easier problems or better luck. Recursion needs a second experiment: an improvement procedure that is held fixed against one that is allowed to improve itself, compared on their ability to find verified upgrades for problems neither has seen, across several cycles or several brands. Only if the self-modifying procedure finds better upgrades, faster and more cheaply, has the system shown that it is getting better at getting better. That experiment needs enough successive improvement opportunities to be meaningful, and will usually take longer than ninety days.
Start where the clock is short, and earn the right to work where it is long.
10 How This Could Fail
If the moat is improvement velocity, it needs to be measured, and three measures capture it. The first is the time from hypothesis to verified result, reported separately for short-clock and long-clock problems so that they are not compared unfairly. The second is the survival rate: the share of changes proposed by agents that beat the current version against a holdout, read alongside how novel those changes were. The third is the full cost per verified dollar of incremental contribution, including failed experiments, incentives, human review, computing and make-goods. Each needs a minimum threshold for what counts as a worthwhile improvement; otherwise a system can look fast simply by promoting trivial changes. These measures extend those in The Moat Is the Loop, which were evidence density, prediction advantage and replication efficiency, by one derivative. They ask not how good the system is, but how fast it is getting better.
They also make the thesis falsifiable. These are the conditions that would refute it.
| Test | Kill condition |
| Self-improvement | The self-improving agent delivers no more contribution per unit of cost than the fixed agent within the pilot |
| Recursion | An improvement procedure allowed to change itself finds no better upgrades, on unseen problems, than one held fixed |
| Proxies | Leading indicators stop predicting contribution measured against holdouts |
| Exploration | The share of novel tests falls, survival rates rise only because proposals grow timid, and no new high-performing ideas emerge |
| Safety | Incidents or reversals rise as autonomy increases, or any hard constraint on permissions is breached |
| Durability | A competitor with the same models but no evidence base or agreed holdouts matches improvement velocity within a year |
Figure 6. Kill tests for self-improving marketing. Each should be agreed before the system is given more autonomy.
The last test deserves emphasis. The code for systems like the Darwin Gödel Machine and Hyperagents is public, and general-purpose tools for improving agents will become widely available. If a competitor using such tools, with no evidence base and no agreed holdouts, can match a system’s improvement velocity, then the advantage was never in the loop. The claim of this essay is that it cannot, because the scarce input is not the method of improvement but the means of verifying it.
Even if every test is passed, there is a ceiling, and it should be stated plainly. Customers adapt to tactics. Responses have a noise floor below which no system can see. Lift has limits, and some brands have too few customers to run many experiments. Improvement will slow as the obvious gains are taken. That does not defeat the thesis. A system does not need to accelerate forever to win; it needs to keep compounding, and to learn faster or more cheaply than the alternatives. This is bounded self-improvement: not a runaway intelligence, but a business that becomes measurably better at becoming better, for as long as its evaluator stays honest and its exploration stays protected.
- • •
The old martech moat was the cost of leaving. The new one is the value of staying. With self-improving agents, that value should keep rising, and rise faster than anywhere else, because the scarce asset is not the agent that proposes the next improvement. It is the system that can prove it.