There is a large difference between asking an AI system to complete a task and building a workflow that can reliably accomplish meaningful work.
The first is increasingly easy.
The second is where things get interesting.
Over the course of developing Divergent Echoes, we have used increasingly agentic workflows across an unusually broad range of disciplines: software development, computer interaction, visual production, game design, research, analysis, documentation, project planning, quality review, and creative iteration.
A game is a particularly unforgiving place to experiment with these systems.
Code has to work. Interfaces have to feel right. Art has to belong to the same world. Mechanics designed months apart have to remain compatible. A change in one system can quietly affect six others. Creative decisions are subjective, but implementation details frequently are not. And unlike a short demonstration or isolated automation, the project has to retain its coherence across thousands of decisions made over a long period of time.
That makes game development a surprisingly useful laboratory for understanding what agentic work can—and cannot—do for a business.
The lessons extend far beyond software.
For marketing and CRM teams in particular, they point toward a future that is more interesting than simply producing content faster.
The real opportunity is to build teams and systems capable of understanding more context, maintaining more experiences, learning from more signals, and operating at a level of personalization that would previously have been too expensive or complicated to sustain.
From prompts to systems
Much of the early conversation around generative AI focused on prompts.
Write an email.
Summarize this document.
Generate an image.
Explain this code.
Those are useful capabilities, but they still treat AI primarily as an interface: a person asks for something, the model responds, and the interaction ends.
Agentic workflows move the boundary.
Instead of asking a model to produce a single answer, we can give systems goals, tools, context, constraints, and methods for evaluating their own work. An agent can inspect an existing implementation before changing it. It can compare multiple sources of information. It can use a browser or development environment. It can execute code, evaluate the result, identify a problem, modify its approach, and report what happened.
More importantly, multiple specialized agents or processes can participate in the same body of work.
That changes the human role.
The question becomes less:
“What should I ask the AI?”
and more:
“How should this work be organized so that people and agents can perform it reliably together?”
That distinction has shaped much of our development process.
Code: the difference between generation and ownership
Software development provides one of the clearest examples.
Generating code is easy.
Maintaining a real software system is not.
An agent working on an established application needs to understand what already exists, distinguish intentional architecture from accidental complexity, preserve unrelated functionality, recognize dependencies, and verify that its change actually solved the problem.
That requires context.
It also requires restraint.
A technically valid solution is not necessarily the correct solution. A new dependency may solve a problem elegantly while creating unnecessary long-term maintenance. A sweeping refactor may be theoretically cleaner while introducing enormous regression risk. A component that works perfectly on a desktop may fail completely on a mobile device.
The best agentic development workflows therefore look much less like unrestricted code generation and much more like good engineering management.
Inspect first.
Understand the system.
Define the task.
Make the smallest appropriate intervention.
Test the result.
Document what changed.
Escalate uncertainty instead of quietly converting assumptions into architecture.
For a marketing organization, the parallel is direct.
A CRM program has its own architecture: audiences, data sources, fields, automations, integrations, templates, business rules, consent states, lifecycle stages, and institutional history.
An intelligent system should not simply generate another campaign.
It should understand the environment into which that campaign is being introduced.
Computer use: agents can increasingly operate the tools, not just discuss them
Another important shift comes from computer-use capabilities.
Traditional automation works beautifully when an API exists, the inputs are predictable, and the process is explicitly defined.
Business reality is rarely that clean.
Teams live in dashboards, administration panels, creative tools, spreadsheets, analytics platforms, ticketing systems, proprietary applications, and interfaces designed for humans rather than machines.
Agents capable of interacting with those environments create a new layer of possibility.
During development, this means an agent can move beyond telling someone how to perform a task. It can inspect the environment, make changes, observe what happened, and incorporate the result into its next decision.
That does not mean every interface should suddenly be operated autonomously.
It means the boundary between “knowledge work” and “execution work” is becoming less rigid.
For CRM and marketing teams, that could eventually mean workflows in which an agent does not merely recommend that a segment be investigated. It can gather the relevant evidence, inspect the configuration, compare related programs, assemble a proposed change, and present that proposal for human approval.
The value is not just fewer clicks.
It is continuity between analysis and action.
Ideation: breadth becomes cheap, judgment becomes valuable
Creative work presents almost the opposite challenge.
An agent can generate enormous numbers of possibilities.
Names. Concepts. headlines. mechanics. layouts. campaigns. visual directions. alternative explanations. customer journeys.
The bottleneck quickly stops being ideation.
It becomes judgment.
Game development makes this obvious because a plausible idea can still be completely wrong for the game.
A mechanic may be clever but conflict with the pacing.
An image may be beautiful but belong to another visual world.
A line of dialogue may be well written while misunderstanding the character.
A proposed feature may solve a problem that does not actually exist.
Agentic creative systems work best when they have enough context to understand not merely what is possible, but what belongs.
That means giving them durable creative principles, examples, constraints, prior decisions, and criteria for evaluation.
Marketing organizations have always faced the same problem at scale.
Brand guidelines traditionally tell a team how something should look.
A deeper agentic knowledge system can begin to capture why decisions were made, which audiences respond to which approaches, what previous experiments taught us, what language creates legal or regulatory concerns, how different channels relate to one another, and where the brand intentionally allows flexibility.
The result is not infinite automated creativity.
It is more useful creative range inside a better-understood system.
Autonomy: the goal is not maximum independence
One of the easiest mistakes to make with agents is to treat autonomy as a score.
More autonomous must be better.
Our experience suggests almost the opposite.
The useful question is:
Where does autonomy improve the work, and where should the system deliberately return control to a person?
Some assignments can safely run for long stretches.
Researching implementation options.
Inspecting a codebase.
Cataloging assets.
Comparing structured data.
Testing a defined set of behaviors.
Preparing documentation.
Other decisions deserve much tighter human involvement.
Does this image actually feel like the character?
Is this mechanic fun?
Does this message represent the company correctly?
Is this customer experience appropriate given what we know about the individual?
Would we be comfortable explaining this decision to the person affected by it?
The strongest agentic workflows therefore include intentional boundaries.
Agents need places where they can proceed independently, places where they must verify, and places where human judgment remains authoritative.
This becomes especially important in CRM.
The future of personalization should not be a system that makes increasingly consequential decisions simply because it technically can.
It should be a system that knows which decisions can be safely automated, which require evidence, and which deserve review.
Analysis: more data is useful only if the system knows what question it is answering
Games create enormous amounts of structured information.
Balance data. Player behavior. encounter outcomes. card choices. progression curves. failure states. performance measurements.
Agents are very good at interrogating this kind of material.
They can identify patterns, compare scenarios, summarize anomalies, and explore hypotheses far faster than a person manually reviewing every record.
But analysis without direction can become another form of noise.
A system needs to understand the question.
Are players losing because a particular enemy is overtuned?
Because they misunderstand an intent?
Because one strategy dominates early and then collapses?
Because the interface does not communicate the available choice?
Those are different problems even if the observed metric is identical.
This is another lesson directly applicable to customer data.
Marketing teams frequently possess far more information than they meaningfully use.
Campaign metrics, purchase history, engagement behavior, service interactions, website activity, declared preferences, lifecycle states, channel behavior—the challenge is rarely the complete absence of signals.
The challenge is turning those signals into useful understanding.
Agentic analysis can dramatically reduce the cost of asking deeper questions of that data.
But the organization still needs to decide which questions matter.
Reporting: agents should explain the work, not merely complete it
One of the most valuable habits we developed was asking systems to report what they had done.
Not just:
Completed successfully.
But:
What did you inspect?
What did you change?
Why?
What did you test?
What remains uncertain?
What should happen next?
This matters enormously once agentic work becomes more complex.
If an agent spends an hour investigating a system and only returns the final artifact, much of the value disappears with the session.
A good report turns execution into organizational knowledge.
It allows another person—or another agent—to understand the decision path without repeating the entire investigation.
For marketing operations, this creates possibilities that traditional automation rarely offered.
Imagine a lifecycle optimization process that not only adjusts a program but maintains a readable history of the evidence considered, changes proposed, tests performed, and decisions approved.
That is much closer to an operating system for continuous improvement than a collection of disconnected automations.
Markdown and durable knowledge: context has to survive the conversation
This may be the least glamorous lesson and one of the most important.
Agents forget.
Conversations end.
Models change.
People change.
Projects outlive individual sessions.
The answer is not simply to create longer prompts.
It is to create durable knowledge.
Throughout development, plain-text documentation became a way to preserve decisions, workflows, artistic rules, implementation patterns, discoveries, failed experiments, and operating instructions.
Markdown is hardly sophisticated technology.
That is precisely why it works.
It is readable by people. It is readable by machines. It can live beside the work. It can be versioned. It can be searched. It can be revised. And it does not depend on a particular AI platform remaining unchanged.
This turns documentation from something written after the work into part of the system that makes the work possible.
For many CRM organizations, this is an enormous opportunity.
A surprising amount of institutional knowledge exists only in the head of the person who built an automation three years ago.
Why does this audience exclude those customers?
Why does this field exist?
Why does this campaign wait seven days?
Why are these contacts handled differently?
Why should this integration never be restarted before another process completes?
An agentic organization benefits enormously from making those answers durable.
Not because the documentation is for the AI.
Because it becomes a shared memory for the entire team, human and otherwise.
Artistic control: collaboration does not require surrender
Generative tools create an understandable fear that increased automation will produce sameness.
Sometimes it does.
The way around that problem is not necessarily less AI.
It is stronger direction.
Visual development repeatedly demonstrated that a generated asset can be technically excellent and creatively unusable.
The work improves when the system has a clear visual vocabulary, reference material, compositional constraints, examples of successful work, and a human willing to reject something that does not belong.
The same principle applies to brand.
Personalization without strong creative direction can easily become an enormous machine for generating mediocre variations.
The goal should not be to produce thousands of different messages.
The goal should be to understand customers well enough to choose when variation matters—and then give the system enough creative context to make those variations feel intentional.
Scale should amplify identity rather than erase it.
What this means for marketing and CRM teams
The most exciting application of agentic workflows is not replacing the marketer.
It is expanding the amount of complexity a good team can responsibly manage.
Today, practical limitations often force customer experiences toward the middle.
A team may theoretically want dozens of lifecycle paths but only have the capacity to maintain six.
It may possess rich behavioral data but lack the analyst hours to continuously explore it.
It may want to personalize messaging more deeply but know that every additional variation creates another artifact that someone has to write, test, approve, monitor, update, and eventually retire.
It may have sophisticated automation but depend heavily on one or two people who understand why everything works.
Agentic systems can change those economics.
A well-designed workflow could help a team:
- Examine customer behavior and surface emerging segments that deserve attention.
- Maintain more lifecycle variants without multiplying operational work at the same rate.
- Generate and evaluate creative adaptations within established brand constraints.
- Audit journeys for contradictory rules, stale content, broken assumptions, or neglected branches.
- Compare performance across audiences and suggest investigations rather than simply reporting averages.
- Preserve the reasoning behind program changes.
- Keep documentation synchronized with the system it describes.
- Prepare proposed optimizations for human review.
- Transfer operating knowledge to new team members far more effectively.
- Give experienced marketers more time to focus on strategy, judgment, and the customer rather than mechanical upkeep.
This is where personalization becomes much more interesting.
The industry has spent years talking about sending the right message to the right person at the right time.
The operational reality has always imposed a fourth condition:
as long as the team can realistically build and maintain it.
Agentic workflows have the potential to move that boundary.
The real opportunity is capability transfer
There is also an important implication for consulting.
A consultant who understands these systems should not simply arrive with a collection of AI tools.
The useful work is helping an organization understand where agents belong in its operating model.
What context do they need?
Which tasks are sufficiently defined?
Which systems can they safely access?
What needs human approval?
What evidence should be retained?
How will output be evaluated?
How does knowledge survive between assignments?
What happens when the agent is wrong?
Who owns the workflow after the consultant leaves?
Those are organizational questions as much as technical ones.
And that is ultimately where our development experience becomes valuable to client work.
Building a game forces us to confront agentic systems as they actually behave—not as they appear in a demonstration.
We see where they accelerate work dramatically.
We see where they lose context.
We see where autonomy is useful.
We see where creative judgment still matters.
We see how documentation changes the quality of future work.
We see how better tooling can make an expert dramatically more capable without making expertise irrelevant.
Those lessons can be transferred.
The goal is not simply to use agents on behalf of a client.
It is to help their team become better at working with them.
More capable teams, not merely faster ones
The easiest AI benefits to measure are about speed.
Hours saved.
Assets generated.
Tasks automated.
Those gains are real.
But they may not be the most important ones.
The deeper opportunity is an organization capable of maintaining more knowledge, exploring more possibilities, operating more sophisticated experiences, and learning faster from what happens.
For CRM and marketing teams, that can mean moving beyond broad personalization toward customer experiences that are genuinely more responsive—without creating an operational system too complicated for humans to control.
That requires technology.
It also requires architecture, judgment, documentation, training, and restraint.
Game development has given us an unusually demanding environment in which to learn those lessons.
The product happens to be a game.
The operating lessons are much bigger.