At S4 Capital’s Monks, agents can run autonomously, unsupervised for days

People still set the terms of the work. Before a run starts, staff write the requirements the agents build against and how well they frame the job shapes what comes back. “You can drive some of that with decent prompting,” said ter Haar. From there the agents keep themselves in check. On a multi-day project several of them split the work, test what they code against the original requirements and review each other’s work as they go. They’re “collaborating and checking each other’s work,” said ter Haar. Until a few months ago developers had to build that scaffolding themselves, writing detailed instructions for how agents should divide tasks and pass work back and forth. 

“The current models really just don’t need that anymore,” said ter Haar. “They spin up their own agentic setup.”

The speed of that shift caught him out. Asked on S4’s half-year earnings call in August what surprised him most about AI this year, he pointed to how long agents could now work unsupervised. Once that was settled, the next question for Monks was what else it could hand them. As Ter Haar explained: “Currently this is mostly software based but the direction of travel is clear: with the right context and guardrails the plan is to extend this to other areas of the business.”

In fact, those plans are already underway. Monks is adapting the agentic setup into its workflow for social and performance marketing, said ter Haar. There the agents wait for something to happen and then act on it, he added, whether that’s a trend taking off that needs a brief, a run of comments that need answering or a post that starts to underperform. Engagement data gives them something to check their work against, the way requirements do for the coding agents. 

“We’re trying to sort of deploy that [agentic setup] into social and performance because we believe that we have both enough baseline context as well feedback loops, where we’ve been able to iterate on whether something is performing and resonating,” Ter Haar said.

Client work comes next, though it will move more slowly. Early next year a client (which Ter Haar did not name) will move their social work onto this rebuilt agentic setup running from insights through to community management.

Oversight moves to the client side

What agents can do for clients also depends on each contract and clients want oversight of their own. When ter Haar spoke to one recently, the questions were mostly about the models: which ones are doing the work, where they come from and whether there’s an enterprise contract behind them. Nobody wants “random Chinese models with no sort of MSA in place,” he said. Clients want to see what was “prompted, what was reasoned.”

A handful are going further. They want every agent kept in one registry where they can “turn them off and on,” track what each one spends and rate its work “because some agents might not be doing the work at the level of quality that we want,” said ter Haar of those clients. He added: “Some of the more modern clients have started to understand this agentic pool that’s probably happening in their organization and whether it’s going to be tenable,” he added.

Inside Monks, the agents have raised harder questions about how people work. Getting decent output from them isn’t the hard part anymore, said ter Haar. It’s the workflow. Agents can turn out 50 ideas on brand and on brief in 10 minutes, and most of them hold up. Ask for something bolder and Monks’ marketers get another 50 that look interesting on paper “but maybe only five are OK,” he said. Finding those five can take so long that the “juice isn’t worth the squeeze.” Monks is testing models tuned for creative work but for now it leans “very heavily on smart subject matter experts” to pick the keepers.

That kind of judgement is becoming a bigger part of the job as the work around it speeds up. 

Agile teams used to plan in two week sprints with a standup every morning. Agents that never clock off have squeezed that cycle down. “The two weeks are now two days with agents working nonstop,” ter Haar said. When work moves that fast, people in strategy and creative teams are tempted to let the agent “just go do it” instead of working through the problem alongside it. As a result, Monks is now debating whether to slow its staff down on purpose. 

“We’re now actually having conversations if we should reintroduce some friction,” said ter Haar. “You want people to push and add their point of view and collaborate so it’s interesting to see that conversation happen where people understand this: like this is work that maybe used to take like two weeks. It’s fine to spend two days on it. It’s fine to go through some of the motions.”

Either way, someone pays. Time spent pushing back on agents is time Monks pays people for, and every hour an agent runs on its own tokens. The better the agents get, the harder it is to decide which of the two deserves the money 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’s deciding how much of it goes on tokens. “They are part of what I call the talent budget. That’s the only place for us to put it,” ter Haar added. Teams keep asking for “more tokens, more tokens,” so Monks is testing how many it will need in 2027, while an internal group sorts the jobs that need the best models from those that can run on the cheaper ones.

As ter Haar explained: “We don’t want everybody using the world’s best model for work that can be done with three models below. We have what we internally call our tech council doing all the work on how to get the most bang for our token bucks.” 

What’s left for people

Where all this ends up, ter Haar isn’t sure, and at S4 that’s an uncomfortable thing to admit. The group has cut about 800 jobs over the past year. Headcount stood at roughly 6,150 at the end of June, down from about 6,350 in December. Monks is “focused on how jobs change, and coordinating the work between people and agents,” ter Haar said in an email. For now ter Haar sees people earning their place by making the calls, trusting their gut on which direction to take and feeding the agents the right inputs. No one can say whether it will stay that way. 

“Let’s not mistake autonomy for accountability,” said Fred Schuster, president of Rosie Advisory, a new division of We Are Rosie that helps restructure marketing organizations and workflow. “The more autonomous we make technology, the more desperately we need oversight. Reshaping workflows doesn’t mean replacing humans, which is a mistake too many financial models assume; it means elevating decision making and refocusing people on real-time thinking rather than continuous task performance.”