The Agent Paradox
Why Enterprise Agent Adoption Depends on Human Rhythms, Not Just Model Capabilities
The Problem
Agentic tools, like Claude Code, became an integral part of my workflow as a Model Behaviour specialist at the Browser Company. I used it daily for everything from building evals to analyzing data. However, the more I used agentic tools the more I noticed a core tension in how these tools are currently designed and how they are fundamentally mismatched for human use.
I often found myself trapped between two equally bad states: watching the agent in paralytic captivity as I waited to approve the next step, or context switching to human-suited work only to return and find the agent blocked, waiting for me.
These tools promise to augment human abilities, but instead humans end up serving the agent through continuous monitoring—what Bainbridge (1983) called the 'ironies of automation'.
This violates well established research on human attention and monitoring of autonomous systems (Parasuraman & Manzey, 2010). We know that humans are actually quite bad at monitoring autonomous systems due to natural cognitive limitations. From studies on air traffic controllers to safety drivers in autonomous cars, humans have proven that they are not capable of monitoring systems that appear to randomly alternate between states of autonomy to needing immediate human intervention.
This is especially important for enterprise agentic systems which are typically not created or managed by their end users—knowledge workers with varying familiarity with AI and agentic tools. These knowledge workers may become frustrated by the start/stop workflow that can occur when trying to work with an agent in parallel. Worse, if they become trapped in a state of agent authorization captivity, they may reflexively approve a dangerous action that violates the company’s AI safety policy.
However, from my vantage point the more dire risk is knowledge workers quietly churning off of a secure enterprise agentic platform to a consumer one that provides a better, more human centric experience. If the enterprise agent UX is hostile to human attention patterns, then all the work on security and privacy is rendered moot the second the end users run sensitive queries on a consumer AI tool. ChatGPT is always just a few taps away on their personal phones.
Model capabilities alone will not solve agentic AI adoption. You can have state of the art performance on standard benchmarks and perform technically well on agentic tasks. However, if the workflow burns humans out they will not use it. Current approaches have engineers manually inserting checkpoints based on risk or compliance. What’s missing is the agents learning when the human attention state warrants interruption. The question for the agent needs to be framed not as “do I need permission to do this task”, but as “is this a natural collaboration checkpoint that respects the human’s attentional state?”. For a more human centred agentic workflow, the human must be understood as a collaborator not just part of the loop.
The Attention Problem is a Design Problem
Our attention spans while using digital devices have collapsed significantly in the last 20 years. Dr Gloria Mark’s work in this field has documented that knowledge workers switched applications every 2.5 minutes in 2004. This has been reduced to 47 seconds as of 2023 (Mark, 2023).
This kind of rapid task switching is known as “kinetic attention” and it’s framed as an adaptation to digital tools that allow for rapid context switching between apps, windows and tabs. While it is easier than ever to access information across multiple tools, this kind of rapid context switching has major costs to human attention. Mark’s work has shown that on average it takes 25 minutes for humans to return to a state of deep focus after a context switch.
Furthermore, Mark details how human attention follows internal circadian rhythms. Unlike AI agents, humans do not have the same capacity for attention throughout the day, but rather the day is broken in to periods of attention peaks and valleys. Typically humans will have two peaks of attention where they are capable of deep work—one in the late morning and one in the early evening. She notes that the attention valleys are equally as important as they act as a rest time for the human where they can accomplish more rote tasks and build up attentional capacity for the next peak. These naturally occurring rhythms of human attention are fundamentally mismatched to the current workflow that requires humans to be constantly monitoring the agentic tool. The risk here is that unless enterprise agentic tools are better aligned with the attentional rhythms of humans they are likely to exacerbate the ‘kinetic attention’ problem of using digital tools, leaving the human even more scattered and drained.
This compounds the vigilance decrement problem documented in autonomous systems research (Parasuraman & Manzey, 2010): humans monitoring automation not only lose situation awareness over time, but in agentic workflows, they're doing so while already operating at fragmented 47-second attention spans.
Recent research has begun acknowledging temporal misalignment between humans and AI. A 2024 CHI paper on proactive AI assistants explicitly studied 'interruption cost' and found that timing of AI interventions significantly impacts collaboration quality (Lee et al., 2024). Similarly, researchers at ACM 2024 called for 'rhythm and attunement' between humans and LLMs, noting that current systems ignore fundamental temporal differences (Buschek et al., 2024). However, this work remains nascent and fragmented. Human-in-the-loop (HITL) frameworks like LangGraph, which govern when agents pause for human approval, focus on technical ‘pauseability’ for compliance, not cognitive optimality for humans. The connection to decades of automation monitoring research—vigilance decrement, out-of-the-loop performance, situation awareness—hasn't been made in the AI industry (Parasuraman & Manzey, 2010; Endsley & Kiris, 1995). And critically, no one is proposing this as a modelling problem—something that could be learned through post-training rather than just engineered into workflows.
Enterprise agentic platforms are typically sold to CTOs and IT heads, but adoption depends on whether knowledge workers actually choose to use them. Security, data sovereignty, and compliance value propositions are meaningless if end users route sensitive queries to consumer tools on their personal phones. Human-centric agent design isn't UX polish. It's a prerequisite for the enterprise value proposition to hold.
Building for Human Time
I’ve been working at the intersection of model behaviour and real user pain on a production AI product for the last 18 months. I’ve debugged where prompt engineering hits a ceiling against the current capabilities of models. I've built evals that exposed systemic behavioural issues. I've watched users abandon well-designed features because the workflow burned them out. These problems aren’t just technological; they’re also human.
The path forward requires treating human attention as a first-class design constraint, not an afterthought. That means agents that learn when to interrupt and when to wait, that work with circadian rhythms rather than against them, and that treat the human as a collaborator rather than an approver. The checkpoints, the pacing, the rhythm of handoffs: these should be learned from data about human attention states, not manually engineered for compliance.
We're building intelligence that operates in human time. It should feel like it.

















