AI Agents
The Agentic Growth Loop: What Should Actually Run Without You
Every vendor now sells an autonomous agent. The useful question is narrower: which parts of the optimization loop genuinely should run unattended, which need a human in the middle, and how to tell the difference before you hand over the keys.
The optimization loop has not changed in twenty years. Observe what happened, form a hypothesis, build the change, ship it, measure, repeat. What has changed is that a machine can now do several of those steps competently.
That does not mean it should do all of them. The steps differ enormously in how expensive a mistake is and how well a model performs. Getting this allocation right is the difference between a team that ships weekly and one that spends a quarter cleaning up after an agent.
Grading each step honestly
| Step | Machine is | Cost of a mistake | Who should own it |
|---|---|---|---|
| Observe what changed | Excellent | Low | Agent |
| Form a hypothesis | Good | Low | Agent, reviewed |
| Build the change | Good and fast | Low while it is a draft | Agent, reviewed |
| Decide it goes live | Poor | High | Human |
| Read the result | Excellent | Medium | Agent, reviewed |
The pattern is not about capability. A model reads a funnel report better than most people and drafts a page faster than any of them. It is about consequence. Four of the five steps are reversible; one is not, and that is the one to keep.
The step nobody argues about, and nobody staffs
Watching every funnel step, every campaign and every experiment continuously is exactly the work humans do worst and machines do best. It is also the work that quietly gets dropped first when a team is busy.
- It is genuinely continuous. A drop-off that starts on a Friday gets found on Friday, not at the Monday review.
- It has no blind spots from habit. People check the dashboards they always check. A machine looks at all of them equally.
- Being wrong is cheap. A flagged anomaly that turns out to be nothing costs thirty seconds.
The publish gate is the whole safety model
Almost every serious failure with autonomous marketing tools has the same shape: something changed in public without a person deciding it should. Copy that misstated a price. A budget shift nobody noticed for a week. A test that quietly kept a losing variant live.
One architectural rule prevents nearly all of it. Everything an agent produces is a draft. Publishing is a separate action a person takes. This is not caution for its own sake. It is what lets you give the agent broad latitude everywhere else.
What a real gate looks like 01
The proposal shows what changes
A diff of the live page against the draft, not a description of it. Reviewers approve what they can see.
What a real gate looks like 02
Approving is one deliberate action
If approval is buried three clicks deep, people batch-approve without reading, and you have a gate in name only.
What a real gate looks like 03
Rolling back is as easy as approving
The willingness to ship quickly depends entirely on how cheap it is to undo.
Widen the loop on evidence, not enthusiasm
The right amount of autonomy is not fixed. It should grow where the agent has a record and stay narrow where it does not. That requires actually keeping score.
- Log the approval rate per workflow. A task you approve ninety times out of a hundred is a candidate for a lighter touch. One you approve half the time is not ready.
- Track what you edit before approving. Heavy editing means the output is nearly right, which is different from being right.
- Widen one workflow at a time. Blanket autonomy hides which change caused a problem.
- Keep the gate on anything that spends money. Even with a perfect record. The asymmetry never goes away.
Most teams that do this land somewhere unglamorous and effective: detection fully automatic, drafting automatic with review, publishing and budget always gated. That is not a limitation of the technology. It is a sensible reading of which mistakes you can afford.
Frequently asked questions
- Q: Is a fully autonomous funnel realistic?
- Technically you can build one today. Whether you should is a different question, and for anything touching public claims or spend the answer is currently no. The cost of a bad publish is asymmetric with the time saved by skipping a review that takes a minute.
- Q: Does review defeat the point of automation?
- No, because review is not where the time goes. Noticing the problem, forming the hypothesis and building the change are the expensive parts. Approving a prepared change takes a minute; producing it took a day.
- Q: How do we know the agent is not just confidently wrong?
- Insist on seeing the evidence alongside the recommendation: the numbers it read and the tools it called. A recommendation you cannot trace is one you cannot evaluate, and that is a property of the tool, not of the model.
- Q: Where should a team start?
- Detection. Let agents watch the funnel and surface what changed for a month before letting them build anything. You get value immediately and you learn how much to trust the judgement before any of it is load-bearing.