Thirty-four percent of enterprise teams now run at least one autonomous AI agent in production, up from fourteen percent six months earlier, according to 2026 adoption data compiled by Soku. Gartner projects that forty percent of enterprise applications will embed AI agents by the end of 2026, up from under five percent in 2025 (cited in the same review). Most of that is happening this year, not behind you.
Here is the part nobody is connecting. Marketers have spent the last decade learning to think in growth loops, systems that reinvest their own output to compound. Engineers have spent the last two years learning to build a different kind of loop, the agent loop, where software gathers context, acts, checks its work, and repeats. Those two loops are about to be the same loop. When that happens, a growth loop stops being something your team runs by hand every week and starts being something that improves itself.
This is what loop and harness engineering, borrowed from how AI agents are actually built, does to growth. Below is what changes, where it is already real in marketing, and the part that decides whether it works or quietly burns budget.
Marketers already think in loops
The funnel is the framework most teams still draw on the whiteboard: acquisition, activation, retention, revenue, referral, a straight line from top to bottom. It is a useful way to describe one step. It is a terrible way to describe how a product actually grows.
Reforge made the argument plainly in 2018: growth loops are the new funnels. A loop is a closed system where the output of one cycle becomes the input of the next. New users create content, content ranks and pulls in new users. Customers fund more ads, ads acquire customers who fund more ads. The fastest-growing companies, as Brian Balfour's team documented, run on a handful of these compounding loops, not on a funnel they keep pouring more money into the top of. There are four broad categories: viral, content, paid, and sales.
The reason loops beat funnels is compounding. A funnel asks "how do we get more at the bottom," and the answer is always "add more at the top," more budget, more channels, more tactics. A loop asks "how does one cohort of customers produce the next cohort," and the answer reinvests instead of refills. That is the difference between renting growth and owning it, the same distinction we draw in our work on owning attention instead of renting it.
So marketers already have the first loop. What they have been missing is a way to run it without a human turning the crank.
Engineers just built the second loop
In the AI world, an agent has a precise and unglamorous definition. Anthropic, which builds the models a lot of these systems run on, defines an agent as an LLM autonomously using tools in a loop. Not a chatbot that answers a question and stops. A system that takes a goal, picks the next action, executes it, reads what happened, and goes again until the job is done.
The shape of that loop, in their own description of how their coding agent works, is gather context, take action, verify work, repeat. The verify step is the one most people skip and the one that matters most. As Anthropic puts it, agents that can check and improve their own output are fundamentally more reliable, because they catch mistakes before they compound and self-correct when they drift.
A loop on its own is fragile, though. Left alone, it loses the thread, repeats itself, or wanders off goal. What keeps it useful is the thing built around it: the harness. In November 2025 Anthropic published a piece specifically on effective harnesses for long-running agents. The harness is the whole apparatus that keeps a loop from falling apart over hours of work. It is the tools the loop can call and the context it is fed. It is the guardrails on what it is allowed to do. It is the memory it writes to, so the next cycle starts where the last one ended, and the evaluator that grades the work and sends it back if it fails.
Harness engineering, then, is the discipline of designing that scaffolding well. The model is the engine. The harness is the car. You do not get anywhere useful with an engine sitting on the garage floor.
What happens when you wrap a growth loop in a harness
Now put the two together.
A growth loop has the same anatomy as an agent loop. Observe what is happening in the market and the account. Decide what to change. Act on it. Measure the result. Learn, and feed that learning into the next cycle. For twenty years a team of humans has been the thing closing that loop, in a Monday meeting, on a Friday report, in the gap between noticing a winning ad and actually scaling it.
Wrap the growth loop in a harness and the closing happens continuously instead of weekly. The system observes performance in real time, decides which variant or budget or audience to shift toward, acts, measures, and reinvests the result, all inside the same compounding structure marketers already understand. The growth loop does not change shape. It changes operator. It goes from a growth loop your team runs to a growth loop that runs itself, with your team setting the goal and watching the guardrails.
That is the precise definition the field is converging on. Ahrefs describes agentic marketing as work where an AI system "takes a goal, picks the steps, runs the tools, checks its own output, and keeps going until the job is done." Autonomous marketing is the same idea at the system level: you set the objective and the boundaries, the loop does the work. The interesting question is no longer whether this is possible. It is which of your loops are ready for it.
Where agentic marketing loops are already running
This is not a forecast. Self-improving growth loops are already running in four places, and they map cleanly onto the categories above.
Self-optimizing landing pages. The clearest example, because the loop is so contained. A page generates variants of its own copy, layout, and visuals, sends traffic to them, measures which convert, keeps the winners, and generates new variants to test against them, continuously. Coframe, one of the tools doing this, reports on its own site a 410% conversion lift on Replit's enterprise funnel and typical lifts of 20 to 30 percent across pages. We pull this apart in detail in how a self-optimizing landing page actually works.
Self-optimizing paid media. This one is already in your ad account whether you asked for it or not. Meta Advantage+, Google's AI Max, and Performance Max are loops: they reallocate budget toward what is working, expand audiences, and pause what is not, on their own. Google reports AI Max lifts conversions by about 14 percent at a similar cost per acquisition. Independent platforms like Albert.ai run the same loop across channels. The kill-the-loser, scale-the-winner cycle that a media buyer used to run by hand is now a loop the platform closes by the hour.
Self-improving creative. At Stripe's 2025 Sessions, Mark Zuckerberg described Meta's destination plainly: a business connects its bank account and an objective, and Meta generates 4,000 ad variants and runs the campaign end to end. Creative becomes a generate, test, kill, regenerate loop. The discipline that used to be the bottleneck, producing enough distinct concepts to keep testing, stops being one. What it does not solve is which concepts are worth testing, which is exactly why creative built from the voice of the customer still wins.
Lifecycle and experimentation agents. Behavioral triggers, churn-risk interventions, next-best-message decisions per individual, and always-on testing are all loops an agent can close. The pattern, as Adello describes it, is the same everywhere: read the current state, decide the next action, execute through an API, observe, adjust.
The four share one trait. Each is a loop with a tight feedback signal and a low cost of being wrong on any single cycle. That is the rule for what to automate first.
The harness is the hard part of AI marketing automation
Here is the operator's truth, and it is the part that separates a self-improving loop from an expensive way to make mistakes at scale. The model is the easy part now. The harness is where the work is, and where most teams fail. Across more than €42M in ad spend we have managed for consumer brands over nine years, the pattern is consistent: teams underinvest in the harness and overrate the model.
The numbers say so. One analysis of production agentic systems found that 88 percent of agents fail before reaching production, almost always because the team skipped the unglamorous foundation: clean data, clear goals, and governance over what the agent is allowed to do. The teams that succeed report an average return of 171 percent, but they spent two to three quarters building the harness before they switched on a single agent. The AI was never the constraint.
A working harness for a growth loop needs four things, and none of them are the model.
A goal and a metric the loop cannot game. An agent optimizing for clicks will happily destroy your conversion rate to get them. The objective has to be the real business outcome, defined tightly enough that a machine cannot find a cheap shortcut to it.
Guardrails on the action space. The loop can rewrite a headline without asking. It should not be able to touch your pricing, your brand claims, or your checkout without a human in the path. Define what is in bounds before you hand over the keys.
A verification step that grades the work. This is the part marketers borrow most directly from how agents are built. Anthropic's long-running harnesses use an evaluator that checks each unit of work against a hard threshold and sends it back with feedback if it fails. In marketing terms, that is a holdout group and a definition of a real win, not a vanity number, sitting between the loop and the budget.
A human review cadence, not a human in every decision. The model the field has settled on is roughly 70 percent of execution handled by agents, 30 percent kept human: strategy, positioning, creative direction, and the call on which loops to trust. The practical form of that 30 percent is a weekly review where the system reports what it did, what worked, what it killed, and what it is testing next, and a human ratifies or redirects. It is the same cadence a brand would have with a senior account manager, which is also how we run growth with the brands we work with.
One honest caveat, because the category is loud right now. Most marketing "AI use" today is still a copilot writing first drafts, not an autonomous loop. Fewer than a third of AI-using marketers apply it to genuinely agentic work, by Soku's count. Autonomy is not the same as abdication, and a growth loop without a harness is not self-improving, it is just unsupervised. The teams that win the next two years will not be the ones who hand everything to an agent. They will be the ones who know which loops to wrap in a harness and which to keep in human hands.
Key takeaways
- Two loops are merging. The growth loop marketers already use and the agent loop engineers just built are the same shape. Wrapping one in the other turns a loop your team runs into a loop that runs itself.
- The harness, not the model, is the work. A self-improving growth loop needs a non-gameable goal, guardrails on its actions, a verification step, and a human review cadence. 88 percent of agents fail for lack of that foundation, not for lack of AI.
- It is already real in four places. Self-optimizing landing pages, paid media, creative, and lifecycle are all loops being closed continuously today, not in a forecast.
- Automate the loops with tight feedback and low single-cycle risk first. A headline can self-optimize. Your pricing and brand claims should not.
- Autonomy is not abdication. The 70/30 split holds: agents run execution, humans own strategy, guardrails, and the call on which loops to trust.
Map which of your loops can run themselves
Most brands have three or four growth loops worth compounding and one or two worth wrapping in a harness this quarter. Knowing which is which is a strategy question before it is a tooling question. If you want a growth model that names your actual loops and shows which are ready to start improving themselves, book a strategy call with our team.
