Using AI in marketing is nothing special anymore.
Most teams already use it for copywriting, content review, and cleaning up data.
But there's one thing worth pausing on.
The way we're working right now is still stuck at "using AI well."
This post is my attempt to sort out the next step: what it takes to move to a structure where we hand work over to AI.
The difference between how we use AI today and an 'agent'
Today's AI usage basically follows this structure:
- A human gives instructions
- The AI produces a result
- A human reviews and revises it
This boosts efficiency, but when the human stops, the AI stops too.
An agent takes a different approach.
An agent isn't given a simple task. It's given a 'goal.'
For example,
- not "Evaluate this content,"
- but something closer to "Make good content."
And to achieve that goal, it iterates, revises, and tries again on its own.
This difference goes beyond simple automation. It's a change to the structure of the work itself.
The GPT we built isn't an agent yet
It's easiest to understand with our team's custom GPT, 'Android Josh,' as the example.
Josh does a good job evaluating content.
The criteria are clear and the feedback is high quality.
But it's hard to call it an agent yet.
That's because the current Android Josh is built to perform a single task: "evaluate the content."
In other words, it isn't built to accomplish the higher-level goal of "make good content" on its own.
This is the part we have to think through to move to the next stage.
The thing you absolutely need when building an agent: a 'harness'
Once you start building an agent, one question naturally follows.
"How far should this agent go in judging and acting?"
The answer to that question is the 'harness.'
A harness is often thought of as a device for improving performance, but in practice it plays a more important role than that.
"Defining the behavior that must not happen."
As AI's capabilities expand, so does the range of its judgment.
Along the way, it can naturally expand into areas like
- proposing changes to the product itself
- overhauling the entire strategy
- criticizing the choice of market
These can sometimes be valid insights, but given an agency's role and scope of responsibility, they aren't always appropriate behavior.
That's why we need a harness.
A harness doesn't point the direction. It's a device that makes the boundaries clear.
Why a well-designed agent can become better than a human
An agent with a harness applied goes beyond a simple automation tool.
In certain areas, it can deliver more stable performance than a human.
1) Productivity
Humans are constrained by physical time and energy.
And at a certain point, we decide that something is "good enough."
An agent, on the other hand,
- has no time constraints
- doesn't tire of repetition
- can hold the standard consistently
In that sense, it can secure consistent execution.
2) Work ethic
In marketing work,
attitude is a problem at least as often as ability.
- Pushing back on the client's conditions
- Criticizing the strategy itself
- Focusing on evaluating rather than executing
These are understandable on a human level, but they may not fit an agency's role.
An agent with a harness draws clear lines around these areas.
- It finds the optimal solution within the given conditions
- It doesn't step outside the scope of its role
- It operates with a focus on execution
In that sense, it can maintain a more consistent professionalism.
How to build a harness
A harness isn't just a set of rules.
It's the work of organizing the team's way of thinking and its execution structure.
Broadly, it splits into two parts.
1) Defining identity
- Our team's mission
- The kind of marketing we aim for
This becomes the basis for judgment.
2) Structuring execution
- Defining work types
- Organizing processes
- Standardizing repeatable tasks
- Documenting how we use Spray (our internal tool) and other in-house solutions
These need to be structured into units of skills.
In the end, a harness is
"our team's way of working, turned into a system."
The limits of a harness
A harness is an important tool, but not a permanent answer.
As AI models advance, many of the constraints and structures we define today will most likely be absorbed naturally.
So a harness is less a finished strategy and more a transitional structure.
7. Conclusion: a harness is the 'operating strategy' we need right now
Even so,
building a harness at this point in time is meaningful.
Through this process we can
- actually run agents in practice
- accumulate trial and error
- and organize a structure that's our team's own
A harness isn't perfect.
But at this stage, it's the most realistically executable approach.
And this experience becomes the foundation for moving to the next stage.
