A growth team's real work is rarely the campaign. It is the loop around the campaign: the guess about who will respond, the test that proves or kills it, the argument over which channel earned the conversion, and the budget shifted on Monday morning. Most of that loop still happens in spreadsheets, dashboards, and meetings — slowly, and by hand.
Cohenta is the engine Thyn is building to close that loop. It is an approach to AI marketing automation in which growth runs as a continuous experiment, operated by agents under policy rather than by people refreshing reports. Cohenta is early — in active development, not yet shipped — so this is an account of what it is being built to do.
The dashboard is not the system
A dashboard tells you what happened. It does not decide what to do next, and it never acts. So the dashboard becomes a place where humans go to translate numbers into actions — a manual relay station between observation and decision.
Most marketing tools sit on the observation side of that gap. They report, segment, and visualize, then hand the judgment back to a person. The slow part was never the charting; it was the human in the middle, reconciling channels and re-deciding spend.
The premise behind Cohenta is that the loop itself should be the product. Sensing, deciding, and acting belong inside one system, with people setting the bounds rather than executing each step. That is the difference between a reporting layer and AI-powered marketing automation that actually moves.
Marketing is an experiment that never ends. The work is running the loop, not reading the chart.
The Cohenta loop
Cohenta is designed around a single cycle that repeats continuously. Each turn names the same four steps, and each step is a place where an agent does work that a team would otherwise do by hand.
- Hypothesize. State a testable claim — a segment, an offer, a message, a channel mix — with the metric that would confirm or refute it.
- Experiment. Launch the test as a controlled experiment, with holdouts and exposure rules defined up front rather than reconstructed later.
- Attribute. Read which touch actually produced the signal, separating the channels that earned a conversion from the ones that merely stood nearby.
- Optimize. Shift spend and creative toward what worked, retire what did not, and feed the result back as the next hypothesis.
The fourth step is the first step of the next turn. That closure is the whole point: the system does not stop at a recommendation, it acts on the recommendation and measures the consequence, then proposes the next move. Autonomous growth is what this loop produces when it runs without waiting for a human to copy a number from one tab to another.
AI marketing automation, run by agents under policy
The loop is run by AI marketing agents — components that own a step rather than narrate it. One agent proposes hypotheses from prior outcomes. Another configures and launches experiments. Another resolves attribution. Another reallocates budget within stated limits.
What keeps that from being reckless is policy. Policy is the explicit boundary the agents operate inside: maximum daily spend, brands and claims that may never appear, audiences that are off-limits, the minimum evidence required before a change to live budget. People write the policy; the agents act within it.
This is the inversion Cohenta is built around. Instead of a person deciding every action and a tool merely executing keystrokes, the agents decide and execute, while the person decides the rules. The human moves up a level — from operator to legislator — which is where their judgment actually compounds.
Simulation before spend
Live experiments cost real money and real attention, and some hypotheses are not worth testing on the market. Before Cohenta commits budget, it is being built to simulate — to estimate the likely range of outcomes for a proposed experiment and prune the ones that fail on paper.
Simulation does two things. It narrows the set of live tests to the ones worth running, so spend concentrates where the uncertainty is genuine. And it gives the optimize step a prior: a model of expected behavior to compare the real result against, so a surprising outcome is recognized as surprising.
This is what separates disciplined automation from a system that simply burns budget faster. Speed without a model of expected cost is just an expensive way to be wrong. The simulation layer is the correctness check that earns the right to act quickly.
Content as a system, not a backlog
Experiments need material — messages, variants, landing pages, the creative that gets tested. Treating each of those as a one-off task is how content becomes the bottleneck that stalls the whole loop.
Cohenta is being designed to treat content as a system. Variants are generated against a hypothesis, governed by the same policy that bounds spend, and produced at the rate the experiment loop consumes them. Content intelligence here means the loop is fed continuously, and that what gets produced is shaped by what the attribution step has already learned works.
The aim is not more content. It is content that closes faster — generated to test a specific claim, measured honestly, and replaced the moment a better variant is proven.
Where this is heading
Cohenta is still being built, and that order is deliberate: the loop, the policy, the simulation, and the attribution have to be trustworthy before any of it should touch a live budget. The first milestones are about making each step honest in isolation, then letting the loop run end to end under tight policy. Over time, the same machinery should support more channels, richer experiments, and tighter cycles — the loop staying constant while its turns get faster.
The bet is simple. Among the next generation of AI marketing automation tools, the ones that matter will not be better dashboards. They will be systems that run the experiment for you, under rules you set, and show their work. Cohenta is Thyn's attempt to build one.