1. “LLMs pre-commodify ideas” - Sachin Benny. Fun way to think about staying original in an AI world: front-run an idea before it makes it into the training data.

2. “Why are LLMs Smart?” - Kevin Kelly. Great as always from KK on intelligence and exactly where “intelligence” is coming from when an LLM solves a problem for you.

3. “Post agent companies” - Yoni Rechtman and Justin Woodbridge. It’s weird to think about, but the “value” of agents as “digital labor” will naturally decrease as open source models improve and the unit cost of inference plummets. Giving away the digital labor for free, but using that to build a more traditional network effect or data moat seems like the smartest move long term.

Music

In this post above, Anna raises an interesting analogy. Cooking vs baking. She noted that AI right now is incredible for cooking, taking a picture of your fridge, applying some human judgment, and dynamically riffing on a recipe. But for baking? Not so much. Baking is high-stakes! Once the muffin is in the oven, you can’t go back and change anything. Her argument was, when things go wrong in high-stakes knowledge work, you end up spending more time sifting through AI-generated slop trying to fix it than you would have spent just doing the work yourself.

It’s a smart framing, and on the surface makes a lot of sense. But it also perfectly highlights the fundamental gap between the “chatbot” era we are leaving behind, and the “agentic” world we are moving into.

When we use a chatbot (like ChatGPT or Gemini), we are stuck in a one-shot paradigm. You write a prompt, hit enter, and you pray the model bakes the perfect muffin on the first try. If it doesn’t, the burden of iteration (i.e. fixing it) falls entirely on you. We’ve all experienced the pain of prompting and prompting, and getting worse and worse results.

The opposite of a perfect one-shot is raw, brute-force iteration. Trying, failing, tweaking, and trying again until it’s right. But to understand why we haven’t built systems this way until now, we have to completely re-frame our relationship with the conceptual merit of accuracy, efficiency, and quality.

We are programmed to hate inefficiency. If a human baker ruined five batches of muffins just to figure out the right recipe for the sixth, we’d call them incompetent (not to mention extremely inefficient).

Once you start thinking about it, you see this cultural meme everywhere:

• “Measure twice, cut once”

• “If a job is worth doing, it’s worth doing well”

• “Slow is smooth, and smooth is fast”

• “Do it right, or don’t do it at all”

Every piece of cultural wisdom we’ve inherited as a species is designed to minimise waste because, for humans, failure is was expensive. Generating multiple incorrect, broken versions of a task (on purpose) just to find the single right one feels instinctively like “slop” and waste to us.

But today compute is cheap, humans are expensive, and agents aren’t humans.

An agentic system doesn’t need to get it right on the first try. The agent’s job is to embrace the exact brute-force trial and error that we find so distasteful, and handle it entirely in the background. Whilst you sleep, or do something else, autonomously.

To lean into the baking analogy: an agent can very cheaply bake 5 versions of the muffin, run an automated validation loop on the results, quietly throw out the four failures, and only serve you the one that actually worked. It ruins five batches of muffins so you don’t have to.

This is the structural shift we are all navigating right now. We are moving away from AI as a synchronous, real-time conversational partner, and toward AI as an asynchronous, autonomous machine. It’s a weird, exciting and difficult journey.

AI won’t make us more productive if we rely on a perfect one-shot prompt every time. AI will (only) revolutionise knowledge work if it’s allowed to brute-force the trial and error while we sleep. If you can hand over a high-stakes baking problem at 5 PM, and by the time you wake up, the agent has quietly burned through the slop, course-corrected, and left the perfect result waiting for you.

I get it, it feels instinctively wrong to rely on brute force. It rubs up against all of our our ingrained cultural instincts. But when the friction of iteration drops to zero, throwing compute at the wall until it sticks isn’t slop, it’s just smart.

1. “Apple Wins Consumer AI By Default” - MG Siegler. Apple has been copping it for years about why they are so behind in the AI race. History says they shouldn’t be counted out though. Apple has a long track record of waiting and watching new technological shifts whilst taking their time to ship the best experience for mass-market consumers. See Airpods/iPhone/Apple Watch etc….

2. “The old world is dying” - Jasmine Sun. At this point, there can be no doubt that AI (and the broader macroeconomic environment) means that the traditional pathways for younger people are eroding (or evolving?). Maybe this isn’t such a bad thing?

3. “Of Termites & Tokens” - Tom Critchlow. Great and fun piece about how organisations must (should) adapt to a token-first world.

Music

I get sporadic and random emails about everything from recipes to “self-care”. Even the copy feels confused. “Ask anything — really”. Here’s a push notification I just received a few minutes ago. This all feels just incredibly disjointed.

The ChatGPT brand, whilst household and very well known, feels all over the place. Dave Morin described OpenAI’s overarching brand narrative as “portfolio” the other day on his podcast, and I think that’s pretty spot on. OpenAI is trying to be everything to anyone. A jack of all AI-trades.

If OpenAI’s narrative is “portfolio”, Anthropic’s is probably “fear”, but of course they’re trying to establish “thinking” as another brand pillar too.

Google’s core narrative is probably “infrastructure” and Apple's might be their classic “it just works” (when it actually works…) alluding to local and on-device AI integrated into to all the apps.

Now, the counter to all this is, ChatGPT is probably the most mainstream and household brand name in all of AI, which sure counts for a lot. But the counter-to-this-counter is that ChatGPT’s once-dominant stranglehold on consumer AI seems to be slipping. Is ChatGPT at risk of becoming the Myspace of the AI era?

Some might say that this is just classic segmentation and focus. OpenAI is going after mums and normies (huge market), whilst Anthropic instead is going after knowledge workers and “ambitious people” (smaller but more valuable market per user?). But it’s not that simple IMO. OpenAI is pushing me to make fun selfies and figure out what to make for dinner, but also wants to own the future of agentic software engineering with Codex. OpenAI wants their cake and to eat it too.

Ultimately, OpenAI will have to fight really hard to maintain their share, given this market is just heating up. Apple is ready to launch Siri 2.0 and Gemini is already a first-class citizen of every Android device, of which there are billions. Google and Apple are both going to be very formidable competitors in the race to win the consumer/normie market.

Can OpenAI have it all? Can you win the suburban mums and the coders? Time will tell, but either way, it’s fascinating to see these giant players in the frontier lab space battle it out in the differentiation arena.

There was something unique about music discovery during this period of time. The internet as a distribution mechanism for music was just starting to become truly meaningful. Enough time had passed since Napster, and big bands were blowing up thanks in part at least to platforms like MySpace. It was the era of the “free mp3” and transferring music from your laptop to your iPod with a USB cable. Even the streaming services (Spotify didn’t get to Australia until 2010-ish) were “library-first”. You had to search and tend to your your library like a garden, carefully.

It’s cliche, but there was something magical about stumbling upon your new favourite song in some weird corner of the internet. The serendipity of it all. I’m sure previous generations felt the same way about the record store, but actually for precisely the same reason. You went looking, and you found something. You weren’t served it.

Anyway, during this “golden age” of musical discovery, I was very into the indie music scene. I read (brilliant) independent blogs like Gorilla vs Bear, You Aint No Picasso, I Guess I’m Floating, Tiny Mix Tapes and yes, a lot of Pitchfork too. These blogs ranged in size and popularity, some of the smallest probably had just thousands of daily readers, whereas Pitchfork was obviously a formidable cultural powerhouse at its peak, but all of them felt “intimate” anyway, in a way that has never really been recaptured in the social media/”big internet” era.

I find myself wanting to go back to this version of the internet (or maybe I’m just feeling nostalgic…), but I’m not sure that it’s possible. Most of these blogs have shut down and many of the internal links are almost all dead. This internet is gone, sadly.

So I decided to build a little time-portal for myself, a way to dip back into this era. It’s a list of the 1000 “best” (imo) songs from this era. It has a nice little YouTube player and there are even some categories and themes too.

Listen along if this sort of this is your jam: https://music.jamespember.com/

1. “Jevons Misunderstanding” - Sangeet Paul Choudary. Bit of a grim read, but worth reading. Basically, even though Jevons Paradox means we will likely see increased demand for knowledge work, it doesn’t necessarily mean we’ll need more people to do it.

2. “Two glass boxes” - Ángela Castillo-Gill. Good read on designing AI-powered enterprise products and the new types of personas or users you need to build for —> those who configure the systems and those who consume the system. A totally new type of power-user.

3. “how to enter side doors” - Maja Wilbrink. Great career-advice piece on becoming legible and shooting your shot. The front door is increasingly shut (or just too crowded), side doors is a great way to think about finding future opportunities.

Music

This is something I am thinking a lot about, and a constant topic of conversation in our product team, so as I try to think about how we are adapting to this new reality, I thought it would be helpful for me to just step through each of these points and flesh out the thinking and what it really means.

Let’s go through these one by one.

“Own complete domain areas”

From platform suites, to specialisation and best-of-breed, it feels like our industry changes its collective mind every few years as to what the best way to architect a software stack is.

In the age of AI, it is becoming clear that tools and point solutions are in a very vulnerable position, because AI and coding agents place tremendous negative pressure on the pricing power of these “lighter weight” tools. AI is very capable at building tools, so it’s going to be increasingly difficult to justify higher prices and annual contract expansion. For a SaaS company with any level of ambition, this isn’t a feasible way to operate. However, out of these tool-ashes comes a huge opportunity for software companies to own entire domains. To be more than just a tool. To be a solution that doesn’t just enable the work to be done, but increasingly does the work too.

What’s most interesting here is that sufficiently ambitious product teams can start to flex their product development muscles and expand into more of their domains. By the way, for this piece, my definition of a domain is the combination of raw tooling and the (often unwritten) knowledge of how, when and where to use the tools.

Let me use a real life example from Komo, where I lead product. Historically, our domain has been customer engagement, but more specifically interactive content, a smaller slice of that broader domain. Our customers, which are large global consumer brands (like Live Nation, NBC Universal or the the New York Comic Con) use our product to design and deploy interactive content to their audiences. This can be competitions, quizzes or micro-games, just to name a few examples.

Generally speaking, Komo has been deployed in an enterprise stack alongside existing tools in the Marketing Automation, CDP and CRM categories. We’ve operated within a wider ecosystem of tools and systems that support our domain.

We’ll always play nice (partner, co-sell, integrate) with other software vendors to be sure, but recently, we’ve been seeing a lot of traction with helping customers consolidate more of these “generic martech” capabilities into Komo. Put simply, more of the stack collapsing into one platform.

Why are customers asking for this? For two reasons:

1. Less tools is almost always better. Everyone instinctively knows this.

2. Consolidated tools in a given domain is the only way to reliably deploy agents.

Every big enterprise, including all of our customers, want to deploy agents. They see the vision clearly, they’ve all drank the kool-aid. However, as we have all learned over the past few years, it is just not that easy to get this stuff to work reliably beyond a cool demo. Deploying an agent to run, for example, customer engagement, across 3-5 different software products (CMS, CRM, CDP…), is just not viable. Integration tax, data-sync latency, security concerns, permissions issues and maintaining up to date context are just a handful of the immediate roadblocks here. It’s worth noting that this dilemma exists in practically every domain and software category today.

If you can consolidate all of the tooling (modular capabilities) needed in one place, encode all of the domain expertise (historical learning, what’s been tried and failed, guardrails, playbooks, rules) and then use agents to drive the car, we start to get closer to a true “system of work”-style paradigm.

Five years ago, the prevailing product orthodoxy was to pick a lane and stay in it. It was simply to much effort for a team like ours to expand into new pockets of our domain. “Core business” and all that. Suddenly, there is an opportunity, for strong product teams, with strong customer relationships, to expand quickly into all of the open space in their domain.

“Are fully agentic”

Almost all of the software we use every day, at least until five minutes ago, has been tool-shaped. Tools that enable humans to do work in a more productive manner. The shape of these software tools have been remarkably consistent across multiple generations. Lists, forms, graphs, tables and buttons. Focus-type-point-and-click. I touched on it above, but one really cannot understate just how big of a mental shift it is to think about software as not just tools that help humans do work, but as a system that can do work too.

It’s no surprise to anyone reading this, but we’re of course now at the point where AI is getting fairly good at reasoning, researching, writing, coding and coordination. When these capabilities sit on top of the domain context we talked about above, things start to get very interesting.

This all has obvious implications for pricing power in two distinct ways. Firstly, software can now in theory angle for budget previously tied up in labor costs but secondly, and perhaps most importantly, solutions that do work and achieve outcomes can start to price based on outcomes and the work done. You will no longer want to pay for a tool to help you with project management, you will want to pay for a solution that does your project management.

“Let you interact however you want”

This idea of headless composability has been one of the most common themes of my blog posts over the past year, because I am fairly convinced that it represents a really seismic shift in how we will think about building products in the future.

In the future, product and design teams will consider agents to be as relevant a “user group” to optimise for, as their human counterparts. APIs, CLIs and SDKs are going to be a far bigger priority for product teams than ever before. This will enable customers to use products in totally new ways: from the command line, from a chatbot or via a copilot-style experience in product too.

“Are self-improving”

This is perhaps the hardest one to nail down, largely because there is a massive misunderstanding out there in the community, especially among customers, about how this actually works.

Many buyers assume that the large language models inherently learn on the fly. They expect that every time an end-user chats with a model, it retains that exact conversation and magically gets smarter for the next run. But foundation models don’t improve statically in real-time. If you feed the exact same prompt to a frozen API weights-set, you’re going to get a similar, static level of capability.

This is all trickier than it sounds. For example, we work with a lot of customers in global sport. One of the most common use cases is in-venue activation. But there are a million-and-one little details here which need to be accounted for. Sending some content to fans 5 minutes after their team has just lost an important game and their best player has been injured is a big no-no, but the models won’t inherently understand this domain-specific context. This is the task ahead of product teams today in this era: figuring out how to encode all of the little nuances and details.

The “self-improving” loop doesn’t happen at the model layer; it happens at the application layer. It’s about building the “harnesses” around the model. When a user edits an AI-generated campaign in Komo, or rejects a specific design choice, that telemetry has to be captured, structured, and fed back into a system which attempts to store these “learnings” to better prompt the models next time. This is not out-of-the-box AI capability. True self-improvement in AI-first software means the system is constantly optimising its own prompts, memory retrieval, and domain guardrails based on real-world usage, making the software uniquely better for that specific customer and use case every single day.

“Are strategic systems”

Historically, a platform like Komo has been viewed primarily as a tool for tactical implementation. A marketing team uses us to execute a specific initiative: “We need to launch a quiz for the New York Comic Con by Friday.”

This isn’t unique to Komo tough, this actually goes for almost all enterprise software today. The humans define the strategy in a meeting room somewhere, and the software is just the execution mechanism. The software doesn’t really understand the strategy.

The shift to an agentic world entirely flips this dynamic. We are moving from tactical execution to goal- and strategy-setting.

Instead of telling the software what to build, the team will input a high-level business objective: “We need to increase first-party data capture among Gen Z attendees at Live Nation events by 15% this quarter.” The software then steps up into the strategic role. It analyses historical engagement data, reviews what campaigns worked or failed in the past, designs the optimal interactive experience, schedules the roll out, and monitors performance. The human moves from being the primary builder to being the editor-in-chief. From tool-shaped to strategy-shaped.

“The primary performance vector is the quality and depth of the AI”

The first era of AI copilots, if we are being completely honest, were often extremely dumb. They were largely thin UI skins slapped on top of raw LLM endpoints. We called them “wrappers” back then. Glorified auto-complete bars that lacked any real sense of context, memory, or domain awareness.

In this new era, your core engineering moat isn’t just your UI or your database model. As Des says, “the primary performance vector is the quality and depth of the AI”.

The harness is the product now. We’re already seeing some truly fascinating work here. This world will require moving far beyond stitching together API calls toward building deeply contextual and verticalised systems, and maybe even custom models, similar to what Intercom has done with Fin, or what Canva and Leonardo.ai are doing with visual generation. It means creating sophisticated “cognitive” architectures. The job as a modern product team is no longer just building great software features, the job is building a highly optimised “system” around the AI to maximise its capability to do real work within your specific domain.

These attributes, these new characteristics, form the basis of a new moat. It’s never been more viable to ship internal tools, to rip out overly expensive point solutions and roll tools specifically suited to your needs. This means the bar for what we will consider “premium” software is going to be drastically raised, but the opportunity on the other side may be bigger than any other time in software history.

1. “When everyone can build, what do you actually own?” - Sachee Perera. A few years into the AI boom, it’s now clearer than ever. We have way more software, but it’s still not really any easier to build a software business. Good post outlining why.

2. “What’s going on with Gemini?” - Martin Alderson. So much of the talk in AI is centered around OpenAI and Anthropic, but Google is IMO the most interesting player in the space given its three unique structural advantages (cash, user distribution and novel TPU infra). But, as with most big companies, it’s kind of a sprawling mess right now.

3. “An economist’s case against the AI jobs-pocalypse” - Casey Newton. Good interview and discussion between Casey and a labor economist on what may or may not happen re: jobs due to AI and automation.

Music

if we want to really make a dent in our declining productivity and standard of living, we need more than just sustainable companies that fund a comfortable lifestyle for its employees, we need crazy outliers which not only create ungodly numbers of productive new jobs (just consider how many people google employ), but also develop networks of capital, talent and ideas that lead to clustering effects, where entirely new industries can be developed and nurtured. i’d argue we’ve already seen this in saas in AU (ty canva, safety culture and atlassian!) but now we also need it in robotics, aerospace, transport, clean energy, ai, data centres, manufacturing, biotech, you name it.

so imo the question just becomes, how do you create an environment that does as much as it can to encourage people to attempt to build outliers?

interestingly, i’m not certain that it’s all about tax. after all, i can’t think of one global tech unicorn company from dubai (famously low-tax region), but i can think of a ton (spotify, klarna, lovable, legora...) from sweden (famously high-tax region).

i’d argue this is because great tech “scenes” are ultimately more about culture and less about politics. but. can/should politics help change/nudge the culture?

ultimately the cgt debate is about incentives, vibes and a very delicate risk/reward balancing act. the cgt discount change messes with the risk/reward equation for tech founders, there’s just no question about that. whether one considers it right or wrong, good or bad, isn’t actually the thing. the thing is just what the downstream and second-order effects could be.

1. “Inside Notion” - Brie Wolfson / Camille Ricketts. Like Linear, Notion is a fascinating company to study, because on the surface, it just seemed so vulnerable to AI disruption. It’s “just” docs and tasks after all. Instead, they’re thriving. Lots of good lessons in here.

2. “Reasoning models don’t so much think as navigate” - Robin Sloan. A beautiful little post musing on whether reasoning models are thinking by “thinking harder” i.e. applying more intellectual effort, or just steering or navigating towards the directionally correct “idea-space”. As always, Robin’s posts are short, sweet and crazy thought-provoking.

3. “Sam Altman and Dario Amodei are both walking back their AI jobs apocalypse prophecies as they eye blockbuster IPOs” - Sasha Rogelberg. Turns out, fear-as-marketing was a pretty bad strategy for the frontier labs, at least in terms of mass market sentiment. A lot of people really don’t like AI, so it’s hard not to be cynical about these 180’s. You wonder what they really believe, and say behind closed doors.

Music

1. Whilst the industry is obsessed with agents and agentic-everything, so much of internal enterprise AI adoption is still just LLM chatbots. Engineers are so far ahead of the curve here on what’s possible with coding agents, but almost none of that productivity has trickled down to Marketing, HR or Finance.

2. Almost all AI usage today is single-player. Productivity gains are almost always made in “private spaces”, the Gemini web app or the Claude Code CLI, but I can’t see what you’re talking to Gemini about. I see the output and the artifacts, but I don’t know how you got there. There is still no consensus on what “mission-control” for agents looks like today inside a company. The closest we have is Claude Cowork.

As a weekend project, I decided to build a prototype of how I’d think about solving these problems. I call it Quarterdeck.

What is Quarterdeck?

• One platform for teams in a company to define their processes as Skills.

• A way to wire up Connected Apps and integrations (Slack, Posthog, Github etc).

• Agents that can read Skills and execute code in a safe and secure isolated sandbox.

• Centralised Artifacts so agents can write and read files outside of their sandbox.

• An Automations platform to orchestrate all of this (triggers, crons, scheduling).

How does it work?

1. First, a user defines a Skill. A skill can be any recurring routine or process. The skill has a detailed set of instructions for how a task or process should be completed, what order steps should be completed in, which tools are required and what output or artifact the skill should generate.

   1. A simple example is a Skill could be “Weekly Product Usage Reporting”. The skill could tell the agent to go look at the usage stats in Posthog or Amplitude, reason about the most important stats and trends, compile it into a PDF and send it via Slack every Monday morning to the relevant channels.

2. Skills can be invoked explicitly (i.e. the user tells the agent to use it via Chat), but agents can also figure out which Skills or Tools should be used to complete a task too.

3. Users can setup an Automation which trigger the agent to execute a task or skill, either based on a Trigger or Schedule. “Do this every Monday at 9am” vs “when triggered via API endpoint”.

Each team has their own Workspace, where they can define their own Connected Apps and Skills.

A skill is just a Markdown file.

Users can connect the apps they use using Connected Apps.

Easily setup recurring automation’s using scheduling.

What’s cool is you can combine Skills and Connected Apps

I built it entirely on the Cloudflare stack, using their Agents and Sandbox SDKs. It’s rough, but it does work, and you can see that this is so clearly going to become an important piece of the enterprise-AI puzzle. I’ve written before about this idea that true AI productivity gains are likely driven by a tiny group in most organisations and that they need systems, and I think something like this is the right set of core primitives: Skills, Agents, Connected Apps and Artifacts.

I’m not going to go any further with this because I already have a startup, and there at on of great companies solving for this “properly” (Runtime being a great example out of the latest YC batch) but nevertheless, it was a great way to better understand the emerging tool-chain and think more deeply about the core primitives and concepts in this space which is going to be coming increasingly important.

In most teams I’ve ever seen up close, people reinvent the wheel constantly. Grunt and glue work that could be automated isn’t, and process documentation is outdated and often ignored. Building an “agentic organisation” is a way out of this. It won’t be easy to get there (old habits and all that), but there is a path.

1. “notes on AI, labor and China” - Jasmine Sun. Jasmine’s Substack is must-read for anyone in tech, a refreshing and unique perspective that is neither pro or anti tech. This post is no different, for anyone thinking about AI job loss, east vs west etc.

2. “Why Tool AIs Want to Be Agent AIs” - gwern. There’s a lot of debate in AI circles around “tools vs agents”. Put another way, are we building tools to make humans better/faster/stronger (i.e. the historical precedent for every technological leap) or are we building our replacements (i.e agents). Very good and detailed post. Worth carving out the time for.

3. “The Messy Middle” - Molly Kinder. For some reason this week I’ve been reading a lot about jobs and AI, but I guess it’s a hot topic. AI is exciting, and we’re probably fine in the long-run (Jevons Paradox, new jobs etc), but there will be pain along the way. Sadly.

Music

like most companies, we hold a monthly all hands/town hall meeting, where we all gather and talk through the current state of the business. metrics, wins, challenges, highlights, low-lights etc.

i found it increasingly crazy that our team would spend a few hours every month pulling data, refining the framing and putting it into a deck for the session. reinventing the wheel on a monthly basis, literally.

i’ve long thought that reporting-on-autopilot could surely be possible with AI. all the data and context exists, albeit across 4-5 various places and platforms.

so i built it. it’s a simple set of python scripts that pull data from a few sources (github, posthog, hubspot, slack and some custom metrics spreadsheets) and then uses some skills, a few templates, and importantly an LLM to turn this all into a simple set of outputs: a presentation (just an html file) and a document.

all of this happens in a few minutes and in my experience, does a really good job at scaffolding out 80-90% of what we’d put together anyway. it’s really cool and powerful, but also a glimpse into the future, i think.

the big takeaway, or the big idea on my mind right now is that, for so much knowledge work, the task is just coordination: shuffling data, numbers and information between systems, and distilling this into “context” that can be delivered to other people inside an organisation.

the big promise of AI is just to supercharge all of this coordination work.

1. “Our AI started a cafe in Stockholm” - Andon Labs. A Swedish AI research company let an agent “run” a cafe in Stockholm. “You won’t believe what happened next!”.

2. “The AI Job Apocalypse Is a Complete Fantasy” - David George. Good read on the history of technological revolution, how these changes impact jobs and why this time it may not be any different.

3. “The Best Companies Will Stop Making Software” - Phin Barnes. An interesting take that argues that the “best” software companies in the future may operate more like Nike, a global design studio that controls brand and vision, whilst “outsourcing” the building further down the supply chain.

Music

I just don’t think that’s how things will work in the future. For most companies, massive step changes in productivity will likely be driven by a small number of hyper-productive-and-capable superstars who figure out how to get maximum leverage from the emerging toolchain. Big is no longer better.

80% of people using AI are using Gemini or ChatGPT to polish their emails or generate “docu-slop” that fills the void of a 40-hour work week. Then you’ve got the 20%, the ones building systems and new ways of working that make them infinitely more productive than their colleagues.

The “workslop” crew is just hitting the “make this sound professional” button because they’re (still) playing the game of appearing productive. But from my vantage point, the power law/10x users aren’t worried about the appearance of work, they’re focused on leverage.

The mistake is trying to “fix” the slop. It’s a waste of energy. You will likely never train away the desire to look busy in a system that rewards busyness.

Instead, the only move that matters is identifying that top 20% of power users and figuring out how to institutionalise what they’re doing. If one person has figured out how to use agents, models, skills and a harness to do the work of five people, the goal isn’t to get the other four to use AI. It’s to turn that one person’s workflow into a repeatable system for the whole company.

The workslop is just the sound of the old world dying. Ignore the em-dashes and go find and promote the people building the work-systems of the future.

1. “Where Are You in the Context Supply Chain?” - Educated Guess (aka Bucco Capital). A pretty cold, albeit quite accurate way to think about how AI-driven job disruption may play out on the field. Reminds me of that old quote: “There are only two jobs—making the product and selling the product”. Everything else is cost, a tax.

2. “MCP - It’s Hot, But Will It Win?” - Steven Sinofsky. There’s lots of chatter in our industry about headless, composability and middleware protocols (A2A/MCP etc) and this is a good piece that draws some historical parallels ultimately arguing that middleware rarely plays out as expected/hoped.

3. “AI isn’t coming for your job. It’s coming for your mind” - Tom Slater. I’ve thought about this a lot, the idea that using AI a lot seems to really change the way your brain feels. I feel far more tired at the end of a heavy-AI-use workday. Good piece that argues that we’re going to need to seriously rewire/adapt our brains if using AI all day becomes the norm.

Music

One way I am thinking about the holy trinity that is [agent harness + skills + models] is that it becomes a framework for codifying and actually implementing this ethos.

What consumes time and energy every day and every week?

What problems are blocking growth and success?

Every process, every bottleneck, every strategic pain point should be fodder for your agents, or at least your humans + their agents. Through this lens, everything becomes a “skill issue”.

Let’s use an example. At Komo, one of the most time-consuming things we do is help customers build custom reports based on their data. Very stock standard SaaS customer success stuff.

Historically, this has been actually quite hard and time-consuming.

Export all the data —> sometimes needing a dev to help with a custom export from the database if the product doesn’t support it —> get it into a spreadsheet —> look for some trends and interesting takeaways —> get that into a deck —> try make it look nice …..

But with the power that is models + agents + skills, this becomes quite easy! You just tell the agent what to do, write a reusable manual (SKILL.md) and then all the user has to do is say, “hey agent, for Customer X, use your entry-analysis skill to pull the data, analyse the data and generate the report using the standard template”.

The skill is the unlock here. This is the process documentation that is so often skipped internally, as keeping it up to date manually over time is just too painful. You can also explain tasks in plain English (i.e. “calculate repeat engagement rate”), and the agent will figure out how to write the code to do it.

What used to take days now takes minutes. A previously-nightmarish business process is now easy to execute over-and-over again. Importantly, you can iterate on the process and have it self-improve.

Right now, whenever I run into recurring tasks, blockers or problems, I try and ask myself: is this just a skill issue? Could some custom code and a set of instructions solve for this?

All of this said, there are still tremendous UX problems to solve here. Orchestration, guardrails, tool definition, scheduling/invocation etc. Using agents and skills is still a fairly big technical leap for a lot of business users. Many are trying to solve this problem (see: Claude Cowork etc), and I think we will ultimately triangulate on some very good UX patterns, but we aren’t there quite yet.

1. “The Human Creativity Benchmark” - Contra Labs. Genuinely interesting attempt to build an eval framework for AI in creative work. Their key insight: most benchmarks treat evaluator disagreement as noise to be resolved, but in creative work, disagreement is the signal, that’s “taste”.

2. “AI Water Use Distractions and Lessons for California” - Jay Lund. Everyone has heard the “AI uses lots of water” debate. Turns out, it’s not that simple. This guy actually did the math and found that California data centres use roughly 0.055% of the state’s annual human water use, less than beer production in Central Arizona. TLDR: The discourse is bad.

3. “Atlassian and Twilio Crush the Quarter, Accelerate. Is the SaaSpocalypse Over?” - Jason Lemkin. Two “legacy” SaaS co’s just crushed the quarter which feels very counter to the prevailing narrative (not to mention the stock price). The big question is whether this is a sustainable return to growth or just temporary gain.

Music

1. “AI’s trillion-dollar opportunity: Context graphs” - Jaya Gupta & Ashu Garg. Agents need context to thrive, we know this. This is a good piece on the emerging concept of a “context graph” or “context-as-a-service” as it pertains to SaaS. A good primer on how it affects both legacy players and new startup challengers.

2. “claude code is not making your product better” - Ethan Ding. I’m not sure I buy this entirely, but it’s good to read things you don’t necessarily agree with, of course. This is a good argument against coding agents and is a reminder that more/faster is not always better.

3. “AI populism’s warning shots” - Jasmine Sun. The recent attacks on Sam Altman are (sadly) surely a sign of things to come. It’s clear now that AI is going to cause very significant social upheaval and become a divisive political issue. Jasmine calls it “AI populism”. Worth a read.

Music

You should watch the video and make up your own mind, but I do want to highlight one idea from Aaron in the discussion that I think is worth touching on, and this is what I am deciding to call the “Aaron Levie Theory of Agents”.

It’s about an infinite pool of software engineers. Imagine one person, inside an organisation, who now has access to this infinite pool of smart and capable people, who are never to busy to help them. These “people” can do three distinct things: write code on the fly to solve problems, use existing APIs and services, and use skills or tools to perform work according to set instructions and guidelines.

The code part really matters here, because code is just a very flexible, robust and universal tool for solving knowledge-work problems. Why? Because at the end of the day, everything you would normally do on your computer, is of course, just code. A presentation, a spreadsheet, a website, a form, an application, an internal tool. All of these are just code. This is why coding agents are the first true “killer app” for LLMs.

It’s still early, and in many ways agents are still more promise than reality, but if you imagine what it would mean to give every knowledge worker a potentially infinite pool of software engineers to better do their job, one can’t help but get excited.