AI Is External Executive Function: What Building With Agents Did for My ADHD Brain

What does it mean to call AI "external executive function"?#

Executive function is the set of mental controls that get you from intent to action: initiating a task, breaking it into steps, prioritizing, holding the plan in working memory, and switching contexts without losing the thread. In my own ADHD, the difficult part is often not understanding the work but starting, sequencing, and resuming it. When I say AI agents act as "external executive function," I mean they help with those specific transitions. That framing is my metaphor, not a clinical term or a treatment claim.

I want to be careful with that phrase, because it is easy to turn into a slogan. So let me define the terms honestly, then tell you exactly where AI helped, and where it made things worse.

What changed in my workflow#

• This is a first-person workflow account, not medical advice and not evidence that AI treats ADHD.
• I use an agent to turn fog into a first action, break work into checkable steps, and write a re-entry note before a context switch.
• I do not offload prioritization, correctness, security decisions, or creative taste.
• The same tool that lowers initiation friction can feed hyperfocus, over-scoping, and dependency.
• The useful test is not "did AI do more?" but "did I retain judgment while reducing friction?"

The part of ADHD nobody warns you about#

The stereotype is that ADHD simply means you cannot pay attention. The National Institute of Mental Health's adult ADHD guide is more precise: adult symptoms can include difficulty staying organized, managing time, planning, remembering tasks, and finishing large projects, alongside inattention, hyperactivity, or impulsivity. Not every person has the same symptoms, and a personal resemblance is not a diagnosis.

This article is narrower than a clinical guide. I am describing the parts of my own workday where an AI agent changes the cost of beginning or resuming a task. It does not tell anyone whether to use AI, medication, therapy, coaching, accommodations, or any other support. Those are individual decisions for a person and, where appropriate, a qualified professional.

Here is the shape for me. I have twelve years in software. I can hold a distributed system in my head. The coding was never the hard part. The hard part is the ten seconds before the coding: the vague intent that won't collapse into a first line, the six-item mental list that evaporates the moment I open a terminal, the project I put down on Thursday and cannot re-enter on Monday because the working-memory scaffold is simply gone.

That gap between "I know what I want" and "I have started" is where a lot of my life used to leak out.

Where AI agents remove the friction#

I practice agentic engineering, not vibe coding. I was a working engineer before these tools existed, and I still review every line. But the way I initiate work has genuinely changed, and it maps closely to the friction described above.

Vague intent to first step. The worst moment in my day was never a hard problem. It was an unshaped one. "I should improve onboarding" is not a task; it is a fog. My practical move is to ask for the smallest observable first action: open one file, reproduce one failure, write one acceptance criterion. An agent can do that shrinking on demand. I describe the fog, and it hands back a candidate first move. The candidate can be wrong; reacting to something concrete is still easier for me than generating step one from nothing.

Breaking work into a plan. I hand the agent a messy paragraph and ask for an ordered list of small, checkable pieces. That externalizes exactly the step I stall on. The first plan is not authority; it is an object I can edit. It does not need to be perfect. It needs to exist so I can argue with it instead of staring at a blank file.

Resuming after a context switch. This is the big one for me, and I think it is underrated. Interruptions carry a heavy re-entry cost in my own work. CHADD describes AI as one possible way to externalize information and reduce working-memory load (CHADD). Concretely, when I return to a codebase after a day away, I can ask an agent to reconstruct what changed, what remains unfinished, and what evidence supports the next move. Context that used to live only in my head can now be rebuilt from the repository and notes instead of memory alone.

"Scaffolding" is the word that fits my experience. The agent is not doing my thinking. It is holding the ladder so I can climb.

The engineering angle: my ADHD and the machine want the same thing#

There is a small but real research literature on software engineers with ADHD, and it named things I had felt for years without vocabulary. An ICSE case study found engineers with ADHD tend toward over-engineering on tasks they enjoy and struggle to stop, alongside real strengths in creativity, quick adaptation, and hyperfocus (Liebel, Langlois, and Gama, arXiv). Both halves are me.

The strengths half is why agentic tooling clicks for me specifically. When initiation is externalized, the hyperfocus has somewhere to go immediately, before the window closes. The friction between "want" and "doing" is where my momentum used to die. Remove it and the ADHD engineering brain gets to spend its energy on the part it is actually good at.

The double-edged sword#

I would be lying, and violating my own honesty rule, if I stopped there. The same properties that make AI a good external executive function make it a good enabler of my worst tendencies.

It feeds hyperfocus rabbit holes. An agent never gets tired, never says "this is enough," never gets bored of the tangent. If I am three abstraction layers deep building something nobody asked for, it will cheerfully help me go a fourth layer deeper. The ADHD tendency toward over-engineering on enjoyable tasks (arXiv) meets a tool with infinite patience for exactly that. The result is a beautiful, useless machine and a Tuesday I will not get back. The agent that helps me start is terrible at helping me stop, and stopping is also an executive function.

It quietly over-scopes. Ask for a plan and you often get an ambitious one. For me, a generous ten-step plan is not always help. It can become a new, more elaborate way to feel behind. I have learned to ask explicitly for the smallest viable version and a list of everything being deferred.

There is a real deskilling risk, and it is worse for me. The research on cognitive offloading is not gentle: over-reliance on AI for cognitive tasks is associated with skill deterioration and weakened independent judgment, sometimes called "metacognitive laziness" (Frontiers in Psychology). One line from that work stuck with me: these systems do not offer neutral support, they actively reshape the environment in which coping happens. For an ADHD brain that is already inclined to outsource the effortful parts, that pull is stronger, not weaker. If I let the agent own initiation and judgment, I am not scaffolding a skill, I am letting it atrophy behind a convenient wall.

The line I try to hold: let AI hold the ladder, never let it decide the building is worth climbing. Initiation and working-memory relief, yes. Prioritization and taste, I keep those, because those are the parts that are actually mine.

My offload / keep / stop framework#

This is the boundary I use when deciding whether an agent is helping or merely making activity easier.

Offload to the tool:

• Turn a vague intention into one physical first action.
• Convert a brain dump into a short proposed sequence.
• Summarize the current state of a project before I switch away.
• Draft a re-entry note containing the next action, open question, and relevant file.
• Check a list for forgotten routine steps.

Keep human:

• Decide which problem matters today.
• Choose the trade-off and define what "done" means.
• Verify technical correctness, security, and privacy.
• Decide what deserves to exist creatively.
• Notice when the entire task should be stopped.

Stop and reset when:

• The proposed plan gets larger each time I ask for clarity.
• I cannot explain an output without asking the same model to explain itself.
• I am polishing an abstraction instead of moving the user-visible result.
• The conversation is increasing agitation or avoidance rather than reducing it.

The last list is as important as the first. A tool that helps me start but makes it impossible to stop has solved only half the problem.

Why I'm building an ADHD-friendly way to make music#

This is not abstract for me. I make ambient music as Ancient Prayers, and music production is an executive-function minefield: a hundred tiny decisions, a huge working-memory load holding the whole arrangement in your head, and an infinite number of rabbit holes disguised as "just tweaking the reverb." It is the exact terrain where my ADHD both soars and stalls.

So the tool I am building is designed around the framing in this post. Not "AI makes the music for you," which is the boring, taste-destroying version. Something closer to external executive function for the creative act: collapse the blank-project paralysis into a first move, hold the scaffold of a session so you can walk away and come back without losing the thread, and, crucially, help you stop and ship instead of spiraling. That boundary is also the subject of where AI belongs in my ambient music workflow.

That last constraint is the whole design philosophy. The best version of this tooling makes it easier to start and easier to finish, and leaves the taste, the judgment, and the meaning firmly in human hands. That is my bet, at least, and I am building it to find out.