AI agents working for entire days with no human in sight.
It sounds like science fiction, but this is exactly what’s happening at Monks, the agency behind S4 Capital — and this scenario has already been reality for a while now.
Some work cycles run 36 or even 48 straight hours, with larger projects stretching across multiple days, according to Wesley ter Haar, Monks’ chief AI and revenue officer.
And the most interesting part isn’t just how long these agents keep running — it’s how they organize themselves during that time, splitting tasks, testing their own code, and reviewing each other’s work.
The dynamic of collaborative work that was always a human responsibility is now also happening between machines. 🤖
But what does this actually change inside a company with more than 6,000 employees? That’s where the story gets more complex — and way more interesting.
How AI agents organize themselves without human oversight
When most people think about artificial intelligence in the workplace, they picture a tool that answers questions or generates a piece of text here and there. What S4 Capital is running goes far beyond that. Most of this work, by the way, involves programming the agency’s own software. Handing these tasks over to agents was, in ter Haar’s words, an almost obvious decision since code is precise and full of well-defined rules. Give an agent a solid set of requirements and it can work on them for days on end.
According to the executive, the company has had coding agents running for 36 to 48 hours without making mistakes. And this happens because the agents don’t just execute isolated tasks — they behave like a functional team, with defined roles, distributed responsibilities, and even internal review and validation processes. It’s a structure that closely mirrors the way well-organized human teams typically handle complex projects.
But here’s the thing: people still set the rules of the game. Before a cycle begins, the team writes the requirements the agents will follow, and the way that briefing is structured directly influences the final result. As ter Haar explained, you can guide a big chunk of this process with good prompting. From there, the agents stay on track on their own.
In a multi-day project, several of them split the workload, test what they program against the original requirements, and review each other’s code as they go. They are literally collaborating and checking each other’s work. Up until a few months ago, developers themselves had to build all of this scaffolding by hand, writing detailed instructions on how agents should divide tasks and pass work along.
According to ter Haar, current models simply don’t need that anymore. They assemble their own agent structure automatically. The speed of this shift actually caught the executive off guard. When asked during S4’s August earnings call about what surprised him most about AI this year, he pointed directly to how long agents can already operate without supervision.
What agent autonomy actually means in practice
The word autonomy gets thrown around a lot when it comes to artificial intelligence, but it rarely shows up in a context as concrete as what’s happening at Monks. Cycles of 36 to 48 hours of continuous operation without human intervention aren’t a lab experiment — they’re real projects, with real deliverables, inside a company that depends on those results to keep its operations running.
That completely changes the weight of the conversation. We’re not talking about a future promise but a reality that’s already shaping business decisions, team allocation, and even the way new projects are structured from the start.
With the duration question settled, the next question for Monks became: what else can be handed off to agents? For now, the focus is on software, but the direction is clear. With the right context and proper guardrails, the plan is to extend this approach to other areas of the business. And those plans are actually already in motion.
Monks is adapting this agent framework for social media and performance marketing workflows. In this scenario, agents sit waiting for something to happen and then act — whether it’s a trend that blows up and needs a quick briefing, a wave of comments that needs responses, or a post that starts underperforming. Engagement data serves as a way for them to validate their own work, the same way requirements function for the coding agents.
The logic, according to ter Haar, is that the company has already accumulated enough context and consistent feedback loops to assess whether something is performing and gaining traction. Early next year, a client — whose name wasn’t disclosed — will migrate all of its social media work to this rebuilt framework, covering everything from insight generation to community management.
When oversight shifts to the client side
What agents can do for clients also depends on each contract — and the clients themselves want control over it. When ter Haar recently spoke with one of them, the questions mostly revolved around the models: which ones are doing the work, where they come from, and whether there’s a corporate agreement behind them.
Nobody wants, as he put it, random models with no formal service agreement backing them up. Clients want to see what was requested via prompt and what reasoning the machine applied.
Some are going even further. They want all agents registered in a single place where they can toggle each one on and off, track how much each agent spends, and even evaluate the quality of the work — because, in ter Haar’s words, some agents might not be delivering at the expected quality level. The more forward-thinking clients, he said, have started to realize there’s a whole set of agents probably operating within their own organization, and they’ve begun questioning whether that’s sustainable.
Collaborative work between humans and machines
One of the most fascinating aspects of the model S4 Capital is building is the balance between purely human collaborative work and fully autonomous cycles run by AI agents. In most projects, both worlds coexist — and learning how to make that integration work well is the big operational challenge companies at this stage of adoption are facing.
Internally, agents have raised tougher questions about how people work. Getting a good output from them isn’t the hard part anymore, ter Haar says. The challenge now is the workflow. Agents can produce 50 on-brand, brief-aligned ideas in just 10 minutes, and most of them hold up.
But when you ask for something bolder, Monks’ marketers get another 50 ideas that look interesting on paper, except maybe only five actually work. And finding those five can take so long that, according to him, the effort ends up not being worth it. The company is testing models fine-tuned specifically for creative work, but for now it still leans heavily on human specialists to pick the best ideas.
This kind of judgment is becoming a bigger and bigger part of the job, precisely because everything around it has sped up. Agile teams used to plan in two-week sprints with a quick standup every morning. Agents that never stop have compressed that cycle dramatically. As ter Haar summed it up, two weeks turned into two days with agents working nonstop. 🚀
And when work moves at that speed, there’s a temptation: just let the agent solve everything instead of thinking through the problem alongside it. That’s why Monks is debating something curious — is it worth deliberately reintroducing friction? The idea is to keep people pushing, bringing their perspective, and actively collaborating. It’s healthy to remember that work that used to take two weeks can comfortably take two days, and that it’s worth going through some steps of the process at a measured pace.
Who foots the bill at the end of the day
One way or another, someone pays. The time spent questioning and guiding agents is time Monks compensates its professionals for. And every hour an agent runs on its own consumes tokens, which also have a cost. The better agents get, the harder it becomes to decide which of the two — human or machine — deserves the investment, since both come out of the same budget.
Until recently, ter Haar planned next year’s spending through the lens of a salary budget. Now he has to decide how much of that amount goes to tokens. As he explained, tokens are part of what he calls the talent budget — and that’s the only place where they fit.
Teams keep requesting more and more tokens, so the company is testing how much it will need by 2027. An internal group, nicknamed the tech council, handles sorting out which tasks require the best models and which can run on cheaper options. The idea is simple and smart: there’s no point using the most powerful model in the world for a job that can be done by three more modest ones. It’s basically about squeezing the most value out of every token spent.
What’s left for people in this new landscape
Where all of this ends up, even ter Haar can’t say for sure — and at S4, admitting that is uncomfortable. The group cut around 800 positions over the past year. Headcount stood at roughly 6,150 at the end of June, down from about 6,350 in December. Monks, according to the executive, is focused on how roles evolve and how to coordinate work between people and agents.
For now, he sees people securing their place by making decisions, trusting their own intuition about which direction to go, and feeding agents the right inputs. No one can say for sure whether it will stay that way.
And here’s an important warning. As Fred Schuster, president of Rosie Advisory — a new division of We Are Rosie that helps restructure organizations and marketing workflows — pointed out, we shouldn’t confuse autonomy with accountability. The more autonomous we make the technology, the more desperately we need oversight.
According to him, redesigning workflows doesn’t mean replacing humans, a mistake financial models assume far too often. It means elevating the level of decision-making and repositioning people for real-time thinking instead of continuous task execution.
The S4 Capital example is one of the most concrete and well-documented cases of how AI agents are moving beyond being a tech curiosity to becoming an essential part of how large companies operate. For those still in the early stages of this journey, the Monks case offers a realistic perspective on what’s achievable when AI adoption is taken seriously and integrated in a structured way into existing processes.
This isn’t a transformation that happens overnight — it requires investment, experimentation, constant adjustments, and above all, a genuine willingness to rethink how work is organized. The future of collaborative work between humans and AI agents is still being written — and companies like S4 Capital are among those drafting the first chapters of this story in real time. 🤖
