Cognitive Load In The Age of AI Agents
Mastering the Puppets Can Be Stressful
Hey, fellow Leader 🚀,
I am Artur, and welcome to my weekly newsletter. I am focusing on topics like IT Management, Innovation, and Leadership, with an Entrepreneurial mindset. My goal is to help you navigate the IT corporate landscape. Make better decisions, create awareness, and share real-world stories.
We think AI coding assistants are helping our developers. In reality, they can potentially crush them under an unprecedented wave of cognitive overload.
In the “old school” way of doing things, one of the key ways to protect a team or an individual developer was to ensure they didn’t have to constantly jump from subject to subject unannounced.
For example, managers avoided giving a task to a developer in the morning only to ask them to switch to a different task just a few hours later.
Keeping things steady avoided mental overload, since context switching makes it much harder for a person to pick up an unfinished task later.
Cognitive Load Management
Cognitive load management is one of the key environmental factors I try to protect developers from.
Even if I assign multiple tasks to a developer, those tasks are ideally part of the same project, product, and context. The overload happens when a task forces the developer to switch contexts entirely.
I strongly believe that “vibe coding” (or vibe engineering) will become the norm in a few months or years. Developers will move away from being strictly a Java, C#, or [insert coding language here] developer. Instead become AI-driven developers.
We are heading toward a scenario where technologies are chosen less based on technical requirements and more on non-technical needs.
I understand this is a controversial statement.
However, the true winners in IT are the ones who build products that are actually used, not the ones with the fanciest features in a PowerPoint presentation.
In the near future, it will be normal for an AI-driven developer to work on multiple tasks at the same time. The main issue I want to raise is that those tasks might belong to entirely different projects or products. High cognitive load will become the norm in IT once again.
The Agents Do All The Work
Yes, it’s true. However, it is humans who must verify and validate the work these agents are doing.
Overseeing work across multiple different contexts is a skill traditionally expected of those in leadership positions, but it is very new to developers.
Context switching has always existed in Engineering Teams, but I fear AI will turn this into a serious issue in the near future.
The amount of mental energy required to validate and verify completely different subjects within a single day, at an accelerated pace, is something that is not being discussed.
AI is a new tool, and everyone is excited and playing with it like a new toy. However, this novelty will fade soon, and the enjoyment of seeing AI do a lot of the heavy lifting will start to wear off.
Simply because the heavy lifting will shift from producing code to reviewing and validating AI outputs. AI will not remove the pressure on IT development teams to deliver more features, faster. It will simply accelerate the processes around IT development.
Tomorrow, the new norm will be to deliver more and more features. The pressure to deliver won’t ease, and any buffer that AI generates will quickly vanish into the ether.
Developers will once again be loaded with different tasks, with the added aggravation of continuous context switching.
Cognitive WIP As New Metric
Development capacity might not be the only high-level metric management should use to assess internal IT delivery potential. Processes need to be adjusted to take cognitive load into account, ensuring we avoid mistakes and deliver high-quality work.
Kanban has the notion of “WIP” (Work in Progress) guardrails. It shows how the amount of work relates to the team’s actual size. Triggering a WIP alert means either the team is starting too much new work without closing previous topics, or there is a bottleneck somewhere in the process.
I would advise implementing “Cognitive WIP” metrics within teams to ensure developers don’t work on too many different topics at the same time.
Parallel work is not great for quality. Mistakes happen, and focusing on the task at hand will produce fewer errors, building confidence in both the product and the team’s output.
Switching between radically different, high-effort topics creates a level of fatigue that is difficult to explain.
To convince management, we need to be factual and use concrete metrics to show how trends in user confidence, production stability, and other relevant areas justify a cognitively safe environment for the team.
Developer's Cognitive WIP = [Sum of Project Complexity] / Number of Projects More Focus On The Process
AI is not a magic tool that automatically builds software products faster, better, and stronger. While it can indeed accelerate parts of the software creation process, the core constraints remain the same:
Getting stakeholders to agree on a solution and a common strategy.
Gathering a set of requirements that are validated by end users, or at least receiving some feedback on their validity.
Following an implementation process (from wireframing and prototyping to building and releasing the product) while keeping constant feedback loops at every step of the way.
AI can facilitate code production to the point where prototyping is incredibly easy nowadays.
For example, a click-through prototype can be built within a few hours to gather user feedback at record speed.
This means the focus is shifting from the code generation itself to how well we execute the process of getting the right information from the right people.
Coding used to be a complex task that could take weeks, only for the business to realize later that they had made a mistake with the requirements.
Sometimes, pushing something through the pipeline was done simply because the business was running out of time to properly define requirements and gather feedback.
Now, building easier prototypes straight from wireframes (sometimes “developed” by the business stakeholders themselves without involving IT) is making these feedback loops much more productive.
AI helps with code production. It does not solve stakeholder alignment.
Oversight Is Now A Developer Skill
A developer is like a Master of Puppets now (imagine a Metallica track playing while reading this part).
Agents are doing the work, and the developer is evaluating and assessing the results of that work.
As mentioned earlier in the article, context switching might become a real issue if teams are pressured to deliver multiple projects across multiple business lines or products.
At the bare minimum, developers will need to improve on a skill that was typically linked to people in management: oversight.
This is the ability to gather information in various formats and act on or share information across multiple topics when requested.
This is not something people are born knowing how to do: It comes with territory and practice.
The challenge is no longer just reading code, but challenging the logic and finding the loopholes behind it.
It is about understanding the security implications and common mistakes AI makes, and finding architectural solutions to address them.
It requires constantly challenging outputs and assessing previous decisions to understand if execution needs to pivot to a different path.
Now imagine doing all this in several different contexts at the same time. The thread is real. The solution is to measure cognitive load.
It has been a wild ride since I published the first article on the SoW. Exchanging views with you and hearing feedback on how these articles have been useful is what drives my motivation to write. Leave a comment, subscribe to the SoW, and be part of the community.
If this article resonates with you, or you know someone who might find it useful, just share the link!
That’s it. If you find this post useful, please share it with your friends or colleagues who might be interested in this topic. If you would like to see a different angle, suggest it in the comments or send me a message.
Cheers,
Artur