the blank canvas is the problem

Microsoft says the modern knowledge worker is interrupted every 2 minutes.

Not metaphorically. Not “it feels like every 2 minutes.” Their Work Trend Index telemetry says meetings, emails, and notifications hit people at an average cadence of every 2 minutes. For the top 20% of users by ping volume, that adds up to roughly 275 pings per day.

Gloria Mark at UC Irvine has the other half of the knife: after an interruption, it takes 23 minutes and 15 seconds to return to the original task.

Every 2 minutes, you get hit.

Every hit takes 23 minutes to recover from.

The math is obviously impossible, which is the point. Modern work is not designed for completion. It's designed for reaction.

This is why the “just focus harder” advice always feels insulting. You are not bad at work. You are living inside a machine that was designed to turn your attention into confetti and then sell you another app to sweep it up.

Asana calls this “work about work.” Status updates. Searching for information. Coordinating. Following up. Moving context between tools. Talking about the work so other people can maybe do the work. Their Anatomy of Work reports put this at around 58% of the knowledge worker's day.

58%.

More than half the day is not the work. It's the exhaust around the work.

And then we act surprised that people are reactive.

Buddy, the entire operating system is reactive.

Email is reactive. Slack is reactive. Calendars are reactive. Task managers are reactive. CRM reminders are reactive. Jira tickets are reactive. Even most AI tools are reactive. A blank box waits for you to type something into it, and if you don't know exactly what you want, it sits there with the dead-eyed patience of a vending machine.

This is the part i think most people in AI are missing.

The next wave of AI is not bottlenecked by intelligence. The models are already weirdly good. They can write, summarize, code, search, plan, translate, classify, extract, compare, generate, and occasionally hallucinate a fake citation with the confidence of a McKinsey partner. The capability is there.

The bottleneck is that humans don't know what to delegate.

Not because they're dumb.

Because they were never trained to manage anything except their own inbox.

Let's go backwards for a second, because the history here is funnier than it should be.

In 1903, Frederick Winslow Taylor published Shop Management, one of the foundational texts of scientific management. Taylor was obsessed with separating thinking from doing. The worker executed. Management designed the system. The job of the manager was not to personally inspect every tiny motion on the factory floor. The job was to define the standard, measure the system, and intervene when something deviated.

This became known as management by exception.

The manager should not be buried in routine operations. The manager should only see the exception.

A machine stops. A metric drifts. A worker falls behind the standard. Something breaks the expected pattern. That's when the manager steps in.

I know Taylorism has a deservedly dark reputation. The “Schmidt” pig-iron story at Bethlehem Steel is basically the creation myth of treating humans like programmable meat. Taylor claimed he could raise Schmidt's pig-iron loading from 12.5 tons to 47.5 tons per day by prescribing his movements and rest. The point was efficiency. The cost was humanity. We can say both.

But hidden inside the ugly factory logic is a useful management primitive: humans should not be spending their cognitive energy watching routine flows manually.

They should be designing systems that surface what matters.

Fifteen-ish years later, another productivity myth enters the chat. Ivy Lee walks into Bethlehem Steel and gives Charles M. Schwab a very simple method: at the end of each day, write down the six most important things you need to do tomorrow. Rank them. The next day, start with number one and do not move to number two until number one is finished. Repeat.

The story goes that Schwab later sent Lee a $25,000 check. In 1918 money. Depending on how you calculate it, that's somewhere in the “small house” range today.

The story is probably a little legendary. Productivity stories always are. But the reason it survived is obvious: the core idea is still annoyingly correct.

The scarce resource was never the list.

It was the act of deciding.

For more than a century, productivity systems have been trying to protect one fragile thing: the human ability to decide what matters before the day attacks.

Taylor tried to do it with standards and exceptions. Ivy Lee tried to do it with six ranked tasks. David Allen tried to do it with Getting Things Done and open loops. Every to-do app since 2004 has tried to do it with cleaner lists, prettier checkboxes, kanban boards, recurring reminders, AI summaries, and tiny animations that make you feel briefly like your life is under control.

And yet here we are.

275 pings per day.

58% work about work.

23 minutes to recover from an interruption that arrives every 2 minutes.

This is not a personal discipline problem. This is a systems problem.

A few days ago, Tyler and i onboarded an IT director from UW.

Small research department. Completely grant funded. Budgets under attack. Administrative work everywhere. One teammate to delegate to, when he clearly wished he had two or three more. He found us through LinkedIn, maybe through someone at Pioneer Square Labs, and came in with one of the best one-sentence summaries of the entire market i've heard so far:

“i'm looking for anything that'll do my job for me that i happen to be leading rather than waiting for it to lead me.”

That sentence has been stuck in my head since.

Because that's the dream, right? Not replacing judgment. Not handing the wheel to a stochastic intern with WiFi. Not “AI agents will run the company while i sit on a beach and post founder-mode threads.”

The dream is: i am leading the work instead of the work leading me.

That's it.

That's the whole product category.

We walked him through Doozy. The to-do list that does itself. You make a task, assign it to a Doozy, and the Doozy works with your context across meetings, tools, docs, and apps to get the work done. It can draft follow-ups. It can pull information. It can create tickets. It can monitor recurring responsibilities. It can record meetings. It waits for approval before write actions. It is not magic, but it is very strange the first time you realize your to-do list is no longer just a guilt ledger.

He understood the pain immediately. Budget constrained department, too much administrative work, not enough people. Perfect.

Then he got to the part that actually matters.

He paused and said:

“i think i have to shift the way i'm thinking about this a little bit. i'm used to asking pretty transactional questions... compare these products, find the spec for this type of laptop... question and answer types of things. rather than... it's a little mental model leap.”

There it was.

Not fear. Not distrust. Not “AI is bad.” Not even confusion.

A mental model leap.

This is the thing i've now seen in nearly every customer discovery conversation. People come in thinking they are evaluating a better chatbot. Then somewhere in the demo they realize the object in front of them is closer to an employee, except weird and cheap and fast and needing very explicit management. You can see their face change.

They thought the question was: what can i ask this?

Then they realize the question is: what part of my work can i hand to this?

Those are not the same question.

One is transactional.

The other is managerial.

And most people have never been trained to be managers of their own work.

This is where the “agentic” language gets annoying, because the word is both useful and deeply cursed.

Agentic sounds like something invented in a San Francisco coworking space by someone wearing a $400 black t-shirt. It feels like a word you say right before explaining your “multi-agent orchestration layer” to a room full of people who secretly just want their inbox handled.

But the underlying concept matters.

A chatbot answers.

An agent owns a loop.

That's the difference.

“Summarize this meeting” is not agentic. Useful, sure. But not agentic.

“After every customer discovery call, pull out the objections, the moment their mental model shifted, any activation risks, any feature requests, and compare it against the last 20 calls so i can see the pattern” is agentic.

“Write this email” is not agentic.

“Own follow-up hygiene for every investor, customer, founder, and friend i talk to this month, draft the note when a relationship is about to go cold, and make me approve before anything sends” is agentic.

“Find me a laptop spec” is not agentic.

“Own procurement research for a budget-constrained research department, compare the options against institutional constraints, prepare the justification doc, and flag which purchases will become painful if delayed” is agentic.

The difference is not autonomy. I actually think “autonomy” is the wrong primitive for most real-world AI work right now. People hear autonomy and immediately imagine an agent deleting a production database, emailing the board, and then sending a cheerful “hope this helps!” recap.

No thanks.

The primitive is responsibility.

Can you give the system a bounded responsibility, enough context to understand it, enough tools to act on it, and a review loop so judgment stays with the human?

That is the leap.

And it is much harder than typing a better prompt.

There is a psychology paper from 1993 by Thomas Bateman and J. Michael Crant called “The proactive component of organizational behavior.” They define proactive personality as the tendency to effect environmental change. Proactive people scan for opportunities, show initiative, take action, and persevere until change happens. Reactive people adapt to circumstances rather than reshaping them.

This sounds like corporate personality-test slop until you start watching people use AI agents.

Then it gets uncomfortably real.

Some people look at an agent and immediately see loops.

They say things like:

Can it check this every morning? Can it watch for these changes? Can it draft these every time this happens? Can it compare all of these calls and tell me the pattern? Can it look through my inbox and tell me what i missed? Can it turn meeting commitments into tasks? Can it do this before i remember to ask?

Other people look at the same tool and ask it to summarize a PDF.

Again, this is not an intelligence gap. Some of the smartest people i've talked to still default to the PDF-summary shape. It's just the interface pattern they've been trained on. Search trained us to retrieve. ChatGPT trained us to ask. Email trained us to respond. Slack trained us to twitch. To-do lists trained us to track guilt. None of these trained us to design loops.

My rough guess from customer discovery is that maybe 1-5% of people are naturally proactive enough to extract the full value of agents without being taught.

I don't mean 1-5% can use AI.

Way more people can use AI.

I mean 1-5% instinctively think: “what recurring responsibility can i carve out of my life and hand to a system?”

That number is the whole market problem.

And maybe the whole opportunity.

Because if only a tiny group already thinks this way, the product cannot just be a toolbox for them. It has to teach the rest of the world the behavior.

This is the part that makes Doozy hard.

The product depends on proactivity, but most users arrive reactive.

So the product has to manufacture proactivity.

I wrote in “context transfer is dead” that communication is two activities wearing the same trenchcoat: discussion/planning and context transfer.

I still think that's right, but i think there's a third layer i didn't name clearly enough.

Management.

Not “management” as in one-on-ones and performance reviews and saying “let's circle back” until everyone dies. I mean management in the older, more literal sense: deciding what should happen, defining the standard, noticing deviations, allocating attention, and making sure the right things move without personally pushing every atom yourself.

That layer used to require other humans.

If you wanted to delegate, you needed headcount. Salary. Hiring. Onboarding. Training. Meetings. Another person with their own context gaps, schedule, energy, incentives, and bad days. Delegation had a high fixed cost, so most people didn't do it.

They absorbed.

They became the API endpoint.

A request came in. They transformed it. They sent an output. Repeat. Inbox to brain to spreadsheet. Meeting to brain to Slack. Slack to brain to Notion. Notion to brain to email.

Human middleware.

This is the real job of a shocking number of knowledge workers.

And because delegation was expensive, “i'll just do it myself” was usually rational. Locally rational, anyway. Faster in the moment. Cheaper than explaining. Cleaner than managing someone else.

But “i'll just do it myself” compounded over years turns into a person whose entire job is being interrupted.

Agents change this because they lower the fixed cost of delegation.

Not to zero. That's the mistake. People who think delegation to AI is free have not actually used AI for real work. You still need context. You still need review. You still need taste. You still need to catch the weird failures. You still need to define the loop.

But the fixed cost drops enough that many tasks cross the threshold from “not worth delegating” to “stupid to keep doing manually.”

That's new.

The spreadsheet Jan maintained for our AI consumer meetup is still the clearest example in my own life. 200+ individual edits across February 6th through February 20th. On one day alone, 80+ edits between 7:31 AM and 2:32 PM. RSVP list to LinkedIn to spreadsheet. Copy, paste, clean, format, repeat.

I did the same class of work in 3 prompts with Doozy.

The important part of that story is not “AI made spreadsheet faster.” That's the boring interpretation.

The important part is that a category of work that previously required a competent human to sit inside the loop can now be moved to a review loop.

That's the difference.

Inside the loop: the human performs every step.

Review loop: the system performs the steps, the human checks judgment.

That is what agentic work actually means in practice.

Not robots replacing everyone.

Humans moving one layer up.

Here's the problem: moving one layer up is psychologically unnatural for most people.

If you've spent your entire life being rewarded for execution, delegation can feel like cheating, laziness, loss of control, or all three.

School rewards execution. Here is the assignment. Do it.

Early career work rewards execution. Here is the ticket. Do it.

Most productivity advice rewards execution. Here is the list. Finish it.

The economy has trained millions of smart people to be extremely high-performing reactors.

And high-performing reactors are dangerous because they look successful for a long time.

They are fast. Responsive. Reliable. Useful. The person who always answers. The person who always knows where the doc is. The person who can pull the report. The person who remembers the follow-up. The person who can jump into a messy thread and clean it up.

Companies love these people until they burn out.

Then they replace them with another one.

I say this with love because i was one of them.

I was very reactive for a long time. Not lazy-reactive. Overactive-reactive. The kind of reactive where you're doing a million things, but the day is still mostly steering you. A hackathon appears, chase it. A deadline appears, sprint. A cold email works, follow the new thread. A person opens a door, run through it. A product idea gets exciting at 2am, disappear into it.

That mode built a lot of my life. I don't want to pretend it was useless. It got me into rooms i had no business being in. It helped me create dots. It helped me cold-email my way into things. It made me fast.

But speed is not the same as agency.

Agency is deciding where the speed points.

The shift for me was not some 10-step Notion template. It was three embarrassingly simple things: time blocking, physically showing up every day to an environment surrounded by ambitious people, and working with Tyler instead of operating alone.

Time blocking forced me to decide before the day attacked.

Foundations changed the ambient default. When you sit inside a room of 350 founders building things, you do not need to motivational-podcast yourself into motion. The room does some of the work.

Tyler forced legibility. Another human cannot execute on vibes trapped in your skull. You have to say the thing. Explain the tradeoff. Name the assumption. Decide whether you're tired or whether the idea is actually wrong.

That last piece matters the most for agents.

Collaboration made my thinking legible.

Legible thinking is delegatable thinking.

This is why agents don't just expose whether your tools are connected. They expose whether your thinking has handles.

If you cannot explain the job to a person, you probably cannot explain it to an agent. If you can explain it to an agent, you usually understand it better yourself. This is why my favorite AI interactions are not the ones where the model gives me an answer. It's when it asks the question i should have asked myself 30 minutes ago.

That is the underrated value.

The AI doesn't just execute.

It forces articulation.

This is also why the blank canvas is such a bad product interface for agents.

“What would you like to do?” sounds friendly. It is actually a brutal question.

It puts the full burden of imagination, scoping, context, prioritization, and process design onto the user at the exact moment they have the least mental model for what is possible.

Imagine hiring an assistant and on day one they walk into your office, sit down silently, and say nothing except “how can i help?”

That assistant might be brilliant. Doesn't matter. Bad onboarding.

A good assistant watches. Asks. Suggests. Notices the recurring mess. Learns what you hate doing. Figures out which relationships matter. Sees the Monday report that always causes stress on Sunday night. Points at the pile you have stopped seeing because it became furniture.

This is where i think most AI products are still stuck in 2022.

They are giving people a blank page and calling it empowerment.

But blank pages are not empowering to people who don't know what genre they're writing in.

This is why starter prompts are not enough. A starter prompt is still a prompt. It's a better vending-machine button. Useful, but not the category shift.

The product has to observe enough context to suggest responsibilities.

You had three meetings this week where follow-up was mentioned. Want me to draft them?

This task has repeated every Monday. Want to make it a recurring responsibility?

You keep asking me to pull the same report. Want me to learn the workflow?

This person has appeared in four important threads and you haven't responded in six days. Want a draft?

This meeting generated five likely engineering tasks. Want me to turn them into issues for review?

That's not the agent running wild.

That's the agent creating endowed progress. Starting the card with two stamps already filled in. Giving the user a shape to accept, reject, or edit instead of asking them to invent the whole workflow from scratch.

This is the theological part of product design that i keep coming back to. Infinite capability is not automatically good. Infinite capability without shape becomes paralysis. The job of the product is not to expose infinity. The job of the product is to make the next right delegation obvious.

The funniest thing about all of this is that “proactive AI” sounds futuristic, but the human need underneath it is ancient.

We have always wanted a way to see around corners.

Farmers watched seasons. Merchants kept ledgers. Factory managers watched exceptions. Executives paid for prioritization methods. EAs became the operational memory around leaders. CRMs became institutional memory around customers. Calendars became externalized time. To-do lists became externalized guilt.

Every tool in this lineage tried to answer the same question:

How do i make sure the important thing doesn't disappear just because the urgent thing got loud?

The difference now is that the tool can act.

That's the rupture.

A calendar could remind you about the meeting. It could not attend, transcribe, extract commitments, draft follow-ups, create tasks, and notice that the same objection appeared in three other calls.

A CRM could store the customer. It could not read the last ten interactions, infer that the relationship was cooling, draft the right note, and ask for approval.

A to-do list could remember the task. It could not do the task.

That is why this feels so strange. The categories are collapsing. Memory tools are becoming execution tools. Communication tools are becoming context sources. Task tools are becoming delegation surfaces.

The old question was “where should i write this down?”

The new question is “who owns this now?”

And “who” is starting to include software.

That's a weird sentence. I know.

But it's also the most practical way to think about it.

So where does that leave people?

I don't think the AI divide is going to be “people who use AI” versus “people who don't.” That was maybe true for five minutes. It already feels stale.

The real divide is going to be between reactive AI users and proactive AI managers.

Reactive AI users will ask better questions. Proactive AI managers will create better loops.

Reactive AI users will save 20 minutes. Proactive AI managers will remove recurring categories of work.

Reactive AI users will generate more outputs. Proactive AI managers will decide which outputs should exist in the first place.

Reactive AI users will be prompt engineers. Proactive AI managers will be process designers.

And the most annoying part is that the proactive people will look almost identical from the outside at first. Same role. Same company. Same Slack avatar. Same laptop. But one of them is building operational machinery around themselves while the other is manually pulling levers faster.

That gap compounds quietly.

Then all at once it looks unfair.

I've already seen it with engineering. Tyler and i run 3-6 coding agents in parallel across a 300,000-line codebase with no full-time engineers. The workflow is discuss, plan, implement, review. We do the first two. AI does the second two. Not perfectly. Not magically. But well enough that the whole shape of building changed.

The same pattern shows up everywhere once you see it.

Content. Admin. Financial reports. Customer follow-up. Meeting synthesis. Lead sourcing. The domains look different. The loop is the same.

Discuss. Plan. Implement. Review.

The human parts are not going away. If anything, they're becoming more important. Taste, judgment, prioritization, context, ethics, deciding what matters. All the annoying squishy human stuff that doesn't fit cleanly into a benchmark.

But the execution layer is getting cheaper every month.

When execution gets cheap, the scarce skill becomes knowing what to execute.

That's proactivity.

Not hustle.

Not grindset.

Not waking up at 5am to cold plunge your way into enlightenment.

The ability to look at your own work and see the hidden system inside it. The recurring loop. The exception pattern. The context transfer. The responsibility that could be named, bounded, delegated, reviewed, and improved.

That's the skill.

Most people don't have it yet.

Again, not their fault. The world trained them out of it.

But the tools are going to train it back into them, one good suggestion at a time.

At least, that's the bet.

At the end of the UW call, after walking through the product, the IT director said something that felt like the real close:

“it's just a big mental step to go from transactional to the agentic.”

Then a minute later:

“how do i shift the way i'm thinking about my workflow to include that?”

That's the whole thing.

Not “how do i use this app?”

How do i shift the way i'm thinking about my workflow?

That question is why i can't stop thinking about this category. Because if we do this right, the product is not just doing tasks. It is teaching a new relationship to work.

Less: i am the endpoint everything flows through.

More: i design the system that catches what matters.

Less: i answer whatever gets loud.

More: i decide which loops deserve ownership.

Less: i use AI as a smarter text box.

More: i use AI as operational memory with hands.

That last phrase is weird, but i think it's right.

Operational memory with hands.

A system that remembers enough to help, acts enough to matter, and asks enough that you still own the judgment.

That's not a chatbot.

It's not really a to-do list either, even though that's the shell we're using because humans understand to-do lists.

It's a management layer for people who never had one.

And the reason the first version of this is so hard is because most people don't wake up thinking “i should become a better manager of semi-autonomous software today.”

They wake up with 275 pings.

They wake up already behind.

They wake up inside the reactive machine.

So the product has to meet them there, in the mess, and gently pull them one layer up.

Not by saying “be more proactive.” That's useless.

By showing them the loop they were too busy to see.

come sit by the pond, the water's good → kparsa.com

-parsa