This is why coding broke out early.

Not because software engineering is easy. It isn’t. And the gap between code that merely runs and code that is elegant, robust, secure and maintainable is enormous.

But code still flatters current AI systems because it lives in a comparatively bounded environment. The feedback loop is tighter. The tools are structured. The constraints are clearer. And at a base level, there is still a fundamental test sitting underneath the whole exercise: does it logically and functionally work, or doesn’t it?

That’s a very different environment from most enterprise work.

The lack of self-driving cars in 2026 provides the perfect example of this.

Autonomous systems look far more competent on a sunny day, on wide roads, with clear lane markings and predictable traffic than they do in messy, high-variance environments.

This is why Waymo operates fully-autonomous in San Fransisco, but in wickedly wintery Minneapolis you’ll still find a driver at the wheel when you jump in.

AI works best in tight constraints and firm context.

A clean coding environment is closer to sunny-day driving.

A real business is not.

A real business is icy roads, missing signage, a roundabout no one indicated into, three passengers changing the destination every 20 seconds and someone from legal yelling that the map is out of date.

That matters because too many leaders are generalising from the most agent-friendly, task-ready professional domain to the least agent-friendly ones.

They see AI doing well in software tasks, creating a spreadsheet or slide deck and assume the same level of competence will transfer neatly into brand strategy, campaign operations, financial reconciliation, team and stakeholder management or cross-functional planning.

It won’t. Not cleanly. Not yet.

If you want to get serious about agentic AI, the right question is not “did it impress me?” It is “what would need to be true for this to count as real help rather than clever theatre?”

For agentic AI to be something more than a polished demo, at least three things need to be true.

First, it needs persistent memory. If the system starts every session like it has been hit on the head with a frying pan and needs your context reloaded from scratch, it is not operating like a teammate. It is operating like a very articulate amnesiac.

Second, it needs to create collaborative editable artefacts. Not just polished output in a chat window, but real work products you can inspect, modify, pressure-test and hand around the business.

And I’d extend that standard a step further: the system should not merely create an editable artefact, it should be able to collaborate with you on that artefact in real time, the way a good team mate would. Not “here’s the draft, good luck”, but “let’s work through this together, adjust the assumptions, change the framing, fix the weak spots and keep the context intact while we do it”.

Third, context needs to compound and refine over time. The tenth task should be smoother than the first - not gunk up the context. The AI should refine and improve it’s understanding of your operating environment as time goes by, not force you back to first principles every bloody time.

A critical nuance of contextual refinement here is also knowing what to forget - just like human memory does. We don’t just gather and gather until we’re full. We gather, synthesise, connect and store optimally for retrieval in connected contextual themes - not in god forsaken folders. And we discard from our RAM whats not needed immediately.

Behavioural psychologist actually consider this one of the pre-eminent roles of the human brain; which is to sort and discard and minimise effort and processing.

Anthropic draws a similar line in its own guidance - whilst still maintaining the terminology of agentic. Workflows are predefined orchestration paths. Agents are systems where the model dynamically decides how to proceed. Even more telling, Anthropic explicitly says teams should start with the simplest thing that works, and that in many cases a well-designed workflow or even a strong single-call system is enough.

That should tell you something.

The companies actually building frontier systems are not saying, “turn everything into an agent immediately”. They are saying, “be careful, use the simplest architecture that does the job, and understand the trade-off between flexibility, cost and compounding error”.

What looks autonomous, might look good enough - but good enough is not always operationally durable.

A self-driving car that can drive beautifully in perfect weather is still not ready for a chaotic highway in a hail storm. And an AI that can generate a stunning artefact under controlled conditions is still not ready for the operational quagmire of a real business.

Which brings us to the bit the demo reels always skip.

Quite simply, real knowledge work is not just a chain of tasks.

Real work is missing context, half-correct source data, contradictory stakeholder opinions, weird edge cases, disconnected systems, unclear ownership, shifting briefs and success metrics that no one bothered to define properly in the first place.

That’s why so much agentic rhetoric feels detached from reality.

Not because AI is useless. And not because the category is fake. But because the loudest claims are often coming from people who do not properly understand the texture of the work they are talking about replacing.

There is a difference between automating execution inside a constrained system and replacing human judgment inside a messy social one.

Most white-collar work is not difficult because there are lots of steps. It is difficult because the steps keep changing, the inputs are unreliable, the trade-offs are political, and the definition of “good” is often negotiated in real time.

That is also why so much current “agentic” positioning ends up as theatre. The artefact looks finished. The reasoning trace looks thoughtful. The output speed looks magical. Then the brief changes midway through, the data is wrong, the compliance rule was buried in a PDF no one uploaded, and suddenly the “autonomous worker” needs a full adult supervision model.

Again, that does not mean there is no value here.

It means the value is currently much closer to partial automation than full replacement.

More Iron Man suit than digital employee. More acceleration with review than unattended execution at scale.

And if you miss that distinction, you end up doing two dumb things at once: over-trusting the technology and under-investing in the operating foundations it would need to become genuinely useful later.

If you lead a martech, growth or operations team, the smartest move right now is not pretending true agentic help has fully arrived.

It is preparing your environment so that when the technology does mature, it has somewhere real to plug in.

That starts with a brutally honest audit.

Pick one workflow in your team that people are already calling “agentic”. Then ask whether you can actually provide the conditions that make agentic help viable.

Then act accordingly.

This is the practical opportunity hiding underneath the hype.

Today’s AI, in its closest-to-agentic form, where the domain is bounded, thrives where verification loop is fast and the failure cost is tolerable.

It’s time to prepare the rest of your operating environment for stronger forms of autonomy tomorrow. And stop confusing a polished output with a proven worker.

Because that is the real issue.

The problem isn’t that AI can’t do work. It’s that too many people are mistaking artifact fluency for operational maturity.

And until those are the same thing, “agentic” is still doing a lot of marketing work for capability that hasn’t fully shown up yet.

If you enjoyed this edition, please forward it to a friend who’s looking to level-up their agentic AI and operational game - they’ll love you for it (and I will too) ⏭️ 💌

PS. When you’re ready here’s how I can help you:

Martech House is a hand-picked, private peer group for senior marketing, digital and martech leaders to get sharper thinking, better signals, and more honest conversations than they’ll find at industry events. Applications are now open for the next intake if you want to be part of it (APAC only) - click here to learn more.

Troy Muir | The Ladder