Blog · Perspective

Agents silo. Systems compound.

A dashboard, 2022

Back in 2022, during my time at Adobe, my team and I built something quite simple and scrappy.

We gathered every creative asset we were running in the market and paired each one with its performance data. That was the whole idea. The studio team could finally see what was resonating before deciding what to make next, and the sales and product marketing team could see which messages were landing and adjust how they told the story.

It was essentially just a dashboard. There was no complex intelligence and no automation involved. It was simply a single place where the entire picture finally came together.

Even so, it changed how we worked across every performance channel. The shared view brought visibility and alignment, and it noticeably increased our pace of execution.

What actually made it work

I have reflected on that experience quite a bit this year, and I have landed somewhere that surprised me.

The value was not really the tool. It was that someone could finally see across the whole landscape, and that the insight was not trapped inside whichever team happened to own it.

Before that, nobody could see across, and not because they were not capable or did not care. It is simply not possible for one person to hold that many surfaces in their head at once. Channels, creative variants, audiences, competitive moves, spend pacing, downstream conversion. The number of things you would need to track at the same time to make a genuinely informed decision is more than any of us can carry.

So we organize around that limit. A channel team is really just a way of dividing the problem into pieces small enough for a person to hold. It is why paid sits in one place, lifecycle in another, and creative somewhere else entirely. It is why our org charts look the way they do.

Seventy agents, four hours, the same walls

Here is what is happening now.

In conversations with CMOs this year, I keep hearing a version of the same story. Their team ran a hackathon, sometimes only four hours long, and walked out with dozens of agents. One leader mentioned that her team built more than seventy. She also mentioned, in the same breath, that they had to hire a Director of AI Ops just to put governance around how those agents were being used.

It is genuinely encouraging that teams are moving this quickly. The difficulty is where the agents ended up. Paid built agents for paid. Lifecycle built agents for lifecycle. Creative built agents for creative. Each one is useful inside its lane, and none of them can see the lane beside it.

The walls between the teams did not come down. They were automated. And the underlying problem did not get smaller, it grew, because there are now dozens of new sources of output running at machine speed inside an organization that already struggled to reconcile four dashboards.

This is the trap. Agents remove the human limit that made channel teams necessary in the first place, but if you point them at the org chart you already have, they do not dissolve the seams. They multiply them.

I am not the first person to say this

It is worth being direct about that, because if you follow the analysts you have already encountered a version of this argument.

BCG found that the CMOs furthest along in this transition have restructured around business objectives and customer segments rather than channels, precisely because agents make cross-channel orchestration far more tractable. McKinsey describes hybrid human and agentic teams, where people design and oversee networks of agents, and is explicit that this requires unified identity and data foundations underneath. Reporting on Gartner’s 2026 CMO research agenda frames agentic AI as an organizational design decision that should be settled before tool procurement rather than after it. Google puts it most plainly, noting that AI changes the shape and skill mix of a marketing organization, not simply its size.

So this thesis is not contrarian. It is close to consensus among the people your board already reads.

What I think that consensus underplays is the sequencing. Read quickly, this body of work suggests that you redesign the organization and the systems follow. What I keep seeing in practice is the opposite failure. Teams redesign nothing, deploy agents into the structure they already have, and only later discover that the structure has quietly hardened around them. A hackathon does not feel like a strategic decision while it is happening. It feels like progress. The seams get set before anyone has consciously decided anything.

That is the gap I care about. Not whether the organization should change, because that argument is largely won, but what has to exist underneath so that it can.

What a system looks like instead

The distinction that matters is not agents versus no agents. It is a collection of agents versus a system of them.

A single agent, even a very good one, is really just a faster dashboard or a faster copywriter. It does one job inside one lane. Stack seventy of those together and you have seventy lanes.

A system works differently. It is a set of agents operating from shared context, which is a common view of what is true right now across every surface that all of them read from and write back to.

That shared context is what makes the rest possible. Different agents can observe different surfaces continuously rather than at review cadence, covering paid performance, creative fatigue, competitive activity, answer engines, and community signal. What one agent learns becomes available to the others, so a signal picked up in one surface changes how another interprets what it is seeing. The system can then recommend and execute the next move rather than reporting on the last one, and outcomes feed back in so the recommendations improve over time.

What we built in 2022 could not do any of this, and not because we lacked ambition. The technology simply did not exist. We could look backward at what had already happened and draw better conclusions from it, which was real value and still only half a loop. This finally closes it.

The harder question underneath

The technology question is the easier one. The organizational question is what I find leaders actually wrestling with.

If the constraint that produced channel teams is going away, what should the team look like instead? What does a growth organization optimize for when span of control is no longer the binding limit? Which decisions stay with people, and which ones were only ever human because nobody else could hold enough context to make them?

I do not think anyone has a complete answer yet, including us. What I do notice is that the leaders furthest along are not asking how to add more agents to what they already have. They are asking how the organization itself should be structured for this era, and treating the tooling decision as downstream of that. That strikes me as the right order.

Why we are building MarZen

This is the core belief behind MarZen AI, the company my co-founder Gireesh and I started this year.

We are building the layer that operates across surfaces rather than within any one of them, so that scattered signals across channels, campaigns, and teams become recommendations and workflows that compound growth. It is designed for the organization that is coming, rather than the one that exists because of a limit we no longer have.

We have been building this alongside our early design partners, and they are already seeing genuine value.

If you recognize these challenges within your own organization, please reach out.