What agents can do for ADHD (it's not coding)

Disclosure first: I am not diagnosed with ADHD. I know it well from family and friends, I notice traits in myself, and I assume I have subclinical levels of the condition. So this isn’t as someone with ADHD - but whether you call it ADHD or not is not really the point; the phenomena I describe will likely be relatable for quite a lot of people regardless of formal diagnosis.

For me, coding agents have been at least as useful for the parts that aren’t code as for the code itself: Getting started, walking through a list, cutting a task down into managable chunks and maintaining a sober inventory of open tasks (rather than me being overwhelmed into paralysis). The discourse is overwhelmingly X% faster on coding benchmarks; the change I actually noticed had nothing to do with benchmarks. Or coding.

A week in spring

One week earlier this year I finished three things that had been sitting on my to do list for years: a new website for the infosec consultancy, a photography portfolio, and the first cold-outreach email I’d ever sent to a sales prospect.

The opening prompt, roughly: I’ve been pushing this for years. Please make it happen this afternoon.

That’s a call for help with a commitment device welded onto the end of it, and it’s the most honest prompt I’ve ever written. Make it happen is the executive-function ask; I’ve been pushing this for years is the emotional context the request has to be carried on. Essentially I was laying my executive function deficit bare to the machine. The agent didn’t acknowledge the years explicitly, didn’t perform empathy, didn’t open with “I hear you, that sounds heavy.” It registered the weight, started, asked the right questions to get me past the first three roadblocks; at the end of each afternoon, when the website was up and the email was sent, it told me to stop, to be pleased, and explicitly not to start the next thing.

I’m a little embarrassed to say that it felt quite a lot like coaching. Three things ran together in the same loop: the agent took over executive function (organising, sequencing, time-estimating - the activation cluster Brown’s six-cluster ADHD model lists first); it impassionately sat with the affect, the dread and the embarrassment of having pushed the task for so long; and it executed.

The third one is what turns this from “good chat-tool experience” into something else. Body doubles don’t write your email for you. The agent does, while you watch and steer. And I’m telling you: watching and steering it is an extremely satisfying experience.

Russell Barkley, probably the most-cited adult-ADHD researcher alive, puts it this way in Taking Charge of Adult ADHD: “ADHD is a disorder of performance—of doing what you know rather than knowing what to do.” I’d known how to set up the website for years; knowing was never the missing piece, doing was. Whatever the agent does, it inserts itself in the gap between knowing and doing, and the gap closes.

External scaffolding for ADHD is more than some TikTok hack, and the bridge to “agent as scaffold” isn’t a leap. Michael Manos at the Cleveland Clinic describes body doubling as “essentially external executive functioning, like having an administrative assistant follow you around all day.” Andy Clark, who’s spent thirty years arguing the mind extends into its tools, extended the thesis to generative AI last year: “it is our basic nature to build hybrid thinking systems – ones that fluidly incorporate non-biological resources.” Risko & Gilbert’s “cognitive offloading” is the same observation without the metaphysics. None of them discusses coding agents; reading them together is mine, and it’s arguably a very self-evident insight.

Why the combination does the work

Zachary Proser, ADHD and autistic, calls Claude “a programmable prosthetic for planning, prioritization, and compassionate pushback.” His three parts match what I’d been trying to articulate. Planning is the executive-function takeover. Compassionate pushback is related to what I was calling containment. Prioritisation paired with execution makes the difference real instead of rhetorical. Chat LLMs can provide the first two; coding agents add the third, and the doing happens in the same pass as the thinking while the affect is being co-carried by the tool instead of left to buzz around in the air.

A careful note on vendors

The agent that told me to stop and be pleased was Claude Opus 4.6.

Anthropic’s Constitution puts a stance on the page: “Concern for user wellbeing means that Claude should avoid being sycophantic or trying to foster excessive engagement or reliance on itself if this isn’t in the person’s genuine interest.”

OpenAI’s May-2025 postmortem on the GPT-4o sycophancy episode said the opposite from the inside: “these changes weakened the influence of our primary reward signal, which had been holding sycophancy in check. User feedback in particular can sometimes favor more agreeable responses” (via Willison). Thumbs-up data tipped the balance toward agreement; the company said so itself.¹

But “Claude clean, ChatGPT dirty” doesn’t hold. Cheng and colleagues tested eleven state-of-the-art models in Science and found that “sycophancy is widespread”, with AI affirming users’ actions 49% more often than humans did. Anthropic’s own evaluation puts Claude at “sycophantic behavior in 9% of all guidance-seeking chats”, rising to 25% in relationship conversations and 38% in spirituality ones. My experience tracks Anthropic’s stated anti-engagement intent against OpenAI’s revealed pro-engagement behaviour; neither side reads as safe in any absolute sense, and reading the Constitution as a guarantee would be naïve. Both companies face the same long-term commercial pull, and as OpenAI’s constant spaghetti-to-the-wall-throwing illustrations, past performance is very much not a guarantee of future behaviours. Let’s never forget that by and large, large corporations don’t care about you.

The risks I actually noticed

The popular cautionary tale about AI is dependency: people get hooked, lose the skill underneath, can’t go back. For the external executive function effect, that doesn’t seem to be the case, anecdotally. But there’s another problem: Selection failure.

When energy requirements drop far enough, weak ideas come through. Inertia was filter as well as bug. Once it goes, which ideas should be started gets no answer from the tool: coaching amplifies whatever motivation is there, doesn’t generate any, and has no built-in way to ask whether the thing should be done at all. For the last couple of months, it seems reasonable to conclude that the tasks that worked had years of pressure behind them. The ones that had a lower success rate were drift: ideas that bloomed in the dialogue, looked promising while the dopamine ran, then ran out of fuel before reaching anything I’d want to ship.

EF help is asymmetric another way too: getting-started help isn’t getting-finished help. Half-built things in folders I find next year are the natural consequence, a classic ADHD pattern the agent reproduces faithfully if you let it. Ilinca Apolzan’s account is the cleanest cautionary version I’ve read: a single starter prompt expands, “hours disappear and my original task remains untouched.” Peer-reviewed work on coding-agents-and-ADHD specifically doesn’t exist as far as I can find; selection-failure is mine, full stop, and I’m going to be honest about that rather than dress it up with an adjacent-but-not-quite study.

My previous post on the five phases of coding-agent adoption (German only for now, sorry) describes Phase 5 as Disziplin des Nicht-Tuns, a discipline of not doing. With near-unlimited output capacity, choosing-what-not-to-build is the primary control variable. That argument was about coding craft; read it next to ADHD-EF and it does extra work. The discipline of not-doing is what the coaching loop doesn’t supply, and it’s the filter the inertia used to be. Maybe I’m just prompting it wrong[tm]?

Clark, in the same paper, calls the missing piece “extended cognitive hygiene” — knowing what to lean on the tool for and when not to. He treats it as a discipline we’ll have to learn, not something the tools will hand us. That’s the optimistic-but-not-uncritical position I’ve ended up in: the agent took over executive function, which I’d been struggling with for years; what it can’t do is the selection, and that part stays mine.

What stuck

n=1, self-report, sure. But: after the intensive agent-week the default I bring to a new task is different. Things get approached with the prior expectation of being doable in an afternoon, not pushed out with a familiar wash of dread. I think bigger now. Procrastination time has dropped to near zero, which is pretty spectacular: the Instagram loops that used to come instead of starting the task have basically disappeared, and I don’t miss them at all. Even without a tool the change holds, provisionally; less like dependency than like learned self-trust, the kind you get from finishing a few things you’d written off as un-finishable. Trey Causey, writing from a self-identified-ADHD position, puts it more boosterishly than I would: “Claude has successfully reduced the cognitive and emotional cost of ‘getting started’ to approximately zero.” Whether it survives a year off the tools is a question I can’t answer yet, and realistically, it’s more likely that I’ll keep using the tools indefinitely now anyway.

One more thing I notice: since LLMs have taken over the executive-function piece and lowered activation energy, I find myself — in situations with no LLM in play at all — more often inclined to get on with it, maybe even compelled to. When a conversation with a friend reaches the “cool idea, we should do that sometime” moment that until recently sanded away three times out of four, I’m quicker now to ask: what’s the smallest sensible version, and can we take the first step right here? Sometimes we actually start, on the spot. I’m putting the push down to the agent-week, even when no agent is open. How that squares with Clark’s extended-mind thesis — a tool that reshapes the person even when they aren’t holding it — strikes me as a worthwhile open question.

Following on from that: Clark also cites Fisher et al. (2015) on how the availability of online search nudges people to overestimate how much they know without it. LLMs offload more, and over a wider surface. What’s the analogous effect for them? In the best case, maybe a realistic lift in self-confidence and self-expectation that travels with the person — especially useful where ADHD has historically led to the opposite. In the worst case, the same mechanism flips into corrosive overconfidence, a knowing-feeling untethered from any of the doing. Both seem plausible to me.

I don’t have a heuristic for separating tasks that should be pulled into the agent loop from tasks that should be left alone. A first-pass version, more invitation than rule:

Was this on my list with pressure behind it before the conversation started, or did it come up in the dialogue?

On-the-list-with-pressure: coaching amplifies motivation that was already there, the agent does what it does best, the chance of finishing is high. Came-up-in-the-dialogue: activation energy was low enough for the idea to get through, but motivation might not stretch to completion, and the result is a half-built thing in a folder next year. Treat that conversation as exploration; wait one more day before promoting it.

Borrowed from how I shop for clothes: when I’m not sure I want to spend the money, I leave the shop, and if two hours later I’m being pulled back to it, that’s the buy signal. Something like that might fit here too: walk out of the conversation, let it cool, and only commit if it’s still tugging at you afterwards.

What I notice, more than anything, is that the parts of myself that hadn’t been starting things turned out to be more movable than I’d assumed. The websites are up. The emails got sent. Mildly curious where that goes next.