A prereading companion to the presentation by Piotr — or, if fate and calendar conflicts conspire, a substitute for it.
On March 25th, 2026, Google Research published TurboQuant — a compression algorithm that reduces the working memory of large language models by a factor of six, with zero accuracy loss. No retraining. No fine-tuning. Just better math.
One week earlier, the Kimi team at Moonshot AI published a paper called "Attention Residuals" that rethinks how information flows between layers in a neural network — a mechanism that had gone essentially unchanged since 2015. Their innovation is elegantly simple: instead of every layer receiving an equal, undifferentiated blend of everything that came before it, each layer now queries its predecessors, focusing on what's actually relevant. The result? A model that matches the performance of one trained with 25% more compute. Same quality, roughly 80% of the training budget. A decade-old assumption about how deep networks should work, quietly overturned in a technical report.1
Two breakthroughs in two weeks. Both freely published. Both benefiting everyone — Google, their competitors, open-source developers, your company, mine, the graduate student running experiments on a gaming PC at 2 AM.
I open with these not to impress you with my newsletter reading habits, but to make a point that will save us hours of circular conversation later: the technology is advancing faster than any single organization can track, let alone control. Worrying about picking the "right" model, the "right" framework, the "right" AI-assisted development tool is a losing strategy. It was Copilot, then Cursor, then Claude Code, and now every developer has their favourite clanker2 — and they'll have a different favourite by Q3.
Think of it this way. There are Star Wars fans. There are Star Trek fans. There are Battlestar Galactica devotees. And there are people who will argue, with alarming passion, that one is objectively superior and the others are trash. This is a fallacy. A fun one at parties, but a fallacy nonetheless. The same applies to the AI tooling landscape today.
There is no single right tool. Not yet. Eventually there will be — the way Excel became the spreadsheet, the way Spotify became the jukebox. But that convergence takes time. And when it arrives, the competitive advantage of having figured this out early will be gone. Commoditized. The question is not "which tool?" The question is: are you building the organizational muscle to use any of them well?
Because here's the uncomfortable truth that no vendor pitch will tell you: the technology was ready before you were. The bottleneck is not silicon. It's everything else.
Let me ask you a question: What will you prepare for dinner?
If my eighteen-year-old son asks, the answer involves speed, volume, and at least one source of protein that doesn't require a cutting board. If my vegan wife asks, we're in an entirely different culinary universe. And if my traditional Polish mother-in-law asks — well, she is an extraordinary cook, and the wisest move is to hand her the kitchen and stay out of the way.3
Now add a modifier: "I want to impress the guests who are coming tonight." Suddenly we have a two-by-three answer space — six fundamentally different dinner strategies from a single question, differentiated entirely by who is asking and what they need.
This is how generative AI works. The same prompt, the same model, the same temperature setting — and yet the quality of the output is almost entirely determined by what the model knows about the situation before it starts generating. This "what it knows" is context. And context is the single most underestimated element in enterprise AI adoption.
Without context, a model cannot predict the right next token. But it has to produce something — that's literally its job, mechanically and mathematically. So it produces its best statistical guess, which without grounding in your reality, is mediocre at best and hallucinatory at worst. This is where "AI slop" comes from. Not from bad models. From starved models. A frontier LLM given a bare-bones prompt is like a world-class chef asked to prepare dinner with no knowledge of the guests, no pantry inventory, and no idea whether the kitchen has an oven. They'll produce something. It won't be their best work.
And this is why the recent TurboQuant breakthrough matters beyond raw performance numbers. That 6x compression of the KV cache — the model's working memory during a conversation — means models can hold dramatically more context before hitting hardware limits. In an agentic exchange, where an AI is orchestrating tool calls, reviewing results, and making decisions across multiple steps, the context window fills up fast. Every tool response, every document snippet, every decision log eats into that budget. Compressing the cache doesn't just save money — it lets the model remember more at the moment it needs to act. More context, better decisions. The plumbing matters.
But better plumbing doesn't help if there's nothing flowing through it.
And here's where it gets expensive: there is no such thing as a second first impression. A senior stakeholder tries the company's new AI tool, gets a vague, generic, slightly-off response, and walks away thinking "this doesn't work." They don't come back. They tell their colleagues. The adoption curve flatlines. Not because the technology failed, but because nobody fed it the context it needed to succeed. You get one shot to demonstrate value to the people whose buy-in you need most. Waste it on a context-starved demo and you've lost them — possibly for years.
My belief — and this is the hill I will comfortably stand on — is that the failure to understand how crucial context is, and the organizational shortcuts people take around it, is the primary obstacle to wide AI adoption in the enterprise. Not cost. Not capability. Not security concerns. Context starvation.
Which brings us to the deeply unglamorous topic that underpins everything.
Context for an agentic enterprise can be expressed in two words: knowledge management.
I know. I felt your enthusiasm drain through the screen. Knowledge management — the corporate equivalent of eating your vegetables. Nobody's favourite topic at the leadership offsite. No vendor is going to fly you to a conference in Barcelona to talk about it. And yet it is the compound interest of organizational capability: invisible in the short term, transformative over years, and devastatingly expensive to neglect.
Consider what most enterprises actually have:
A business process documented in a twenty-slide PowerPoint deck, but no BPMN diagram — because creating one is hard and maintaining one is harder. Tribal knowledge that makes a new hire productive in month four instead of day four, locked inside the heads of people who've been around long enough to accumulate it. The "let's meet and talk about this" culture that persists because "I never had time to document this properly — it would take me a full day, so instead I spend thirty minutes explaining it every single week." That last one is my favourite, because the person saying it is simultaneously the bottleneck and the one who feels irreplaceable precisely because of that bottleneck.4
Now picture an agent trying to help with that business process. It has the PowerPoint. Twenty slides of boxes and arrows, probably with a gradient that was fashionable in 2019. No structured data. No decision logic. No exception handling. The agent can describe what it sees on the slides, but it cannot execute the process, because the process as documented is a sketch — not a blueprint. The blueprint lives in Kasia's head, and Kasia is on holiday. Climbing rocks most probably.
Knowledge management is like compound interest: the payback is far from immediate. There is little reward for doing it well and almost no visible punishment for neglecting it — until you try to make your enterprise agentic and discover that the agents have nothing to work with. And unlike financial debt, knowledge debt is invisible on every balance sheet and dashboard in your company. Nobody reports on it. Nobody is accountable for it. It simply accumulates, silently, until the day you need what was never captured.
You have knowledge debt. Years of it. Decades, probably. Processes that exist only as oral tradition. Decisions documented in email threads that have been archived and forgotten. Expertise that walks out the door every time someone retires or gets a better offer. This debt accrues silently, and it comes due the moment you ask an AI system to operate autonomously within your organization.
The good news: you don't have to pay it all off at once. But you do have to start. And you have to be intentional about it — which is a word I'll keep coming back to.
AI is like the steam engine. Like electricity. Like the personal computer. Each of these technologies was ready for enterprise adoption long before enterprises were ready for them. The delay was rarely technical. It was the business transformation required — the rewiring of processes, habits, incentive structures, and power dynamics.
I am forty-six years old. I know from personal, sometimes painful experience how hard it is to change a habit. To unlearn the way you've done things for "a little while." We are not blank sheets of paper that can be rewritten at will. Every efficiency we've built, every shortcut we've internalised, every "I just know how to do this" — these are neural pathways carved over years of repetition. Asking someone to abandon them is not a training problem. It is a neurological renovation. And renovations, as anyone who has lived through one will tell you, always take longer and cost more than the estimate.
Changing how one person gets things done is tough. A department? Harder. An enterprise? You'd better have a plan, patience, and a deep appreciation for the fact that the human brain — the protein brain — is simultaneously your greatest asset and your biggest obstacle.
Your organization is full of intelligence. Protein-based mastery. Years of experience, pattern recognition, judgment, relationships. The person who can glance at a supply plan and spot the anomaly that the dashboard missed. The project manager who knows that a particular stakeholder needs to hear bad news on Tuesday mornings, never Fridays. The engineer who carries the system architecture in their head because the documentation was last updated three reorganizations ago. You cannot and should not try to replace this intelligence. But you do need to augment it — and augmentation requires change, and change requires that the people doing the changing see something in it for themselves.
However cliché it sounds: you need to think win-win. If AI is implemented as a stick — "run faster for the same reward" — you will get compliance at best and sabotage at worst. If it's implemented as forced token consumption — everyone must use X prompts per week — you'll get your scorecard green and your tokens burned, but the value will be somewhere between negligible and imaginary.
Reward what was impossible before and is now possible. Reward impact, not process. If someone uses AI to compress a three-week analysis into three days, the conversation should be "what will you do with the two and a half weeks you just freed up?" — not "prove to me you used the approved tool in the approved way." The former creates an ally. The latter creates a box-ticker.
Track adoption, yes. It helps you understand reality. It helps you decide where to invest intentionally. But do not mistake the metric for the outcome. A dashboard full of green adoption numbers tells you people are clicking buttons. It tells you nothing about whether those clicks are creating value. And it certainly tells you nothing about whether the person clicking has any intention of coming back tomorrow without being reminded.
While humans are — quite literally — boiling the ocean right now5, it thankfully takes time even with considerable energy input. Turning your company agentic in one go is an attempt to boil your own organizational ocean. I wish you luck, but the odds are roughly equivalent to a tossed coin landing on its edge.
There are pockets of excellence in your organization. Leaders who get it. Teams that are already experimenting. Individuals who've built their own workflows with the tools available. Your job is not to launch a grand transformation programme with a steering committee and a Gantt chart that stretches to 2028. Your job is to recognize these pockets, tend to them, kindle that fire — and protect them from the organizational antibodies that will inevitably try to slow them down or standardize them into irrelevance.
Is it a terrible idea to end up with a catalogue of two hundred isolated agents for your company? Yes. Terrible. Is it a great waypoint to pass through on the way to an agentic enterprise? Absolutely. Because those two hundred agents represent two hundred executable intents — encoded context, defined workflows, real problems given a working answer. You need them first. Orchestration comes later, and the orchestration layers of tomorrow will be better than anything available today. Some agents won't survive. That's fine. The ones that do will be battle-tested. And you're nowhere close to two hundred value-added agents catalogued today. Getting there is progress.
Place seeds in many places. They will grow in time and cover the whole plot. The middleware will mature. The integration points will simplify. But the understanding of your own processes, the encoding of your own context — that work is yours to do, and it doesn't depreciate. Every process your team encodes into an agent prompt today is a piece of institutional knowledge that persists regardless of which platform runs it next year.
Warm one cup at a time. Serve it. Get feedback. Repeat.
This connects directly to the knowledge management challenge: you can try to address your entire knowledge debt in a single, Augean effort, or you can grow it incrementally through these same pockets of excellence. Each agent someone builds is an act of knowledge encoding. Each prompt template that works is a piece of process documentation. The incremental approach lets you accumulate knowledge while delivering value, rather than asking the organization to invest in a massive documentation project before anyone sees a return. And those who do the hard, unglamorous work of encoding that knowledge — one process, one template, one agent at a time — deserve to be recognized for it.
Think of it as establishing camps on the way to a summit. You have a direction — a vague north star of what "agentic enterprise" means for you. But you don't plan the entire route from base camp to peak on day one. You plan to the next camp. You acclimatize. You assess conditions. You adjust. Sometimes you discover that the ridge you were aiming for is impassable, and you reroute. The summit is the same, but the path adapts.
Each camp is a delivered benefit. Not a promise. Not a pilot. A working capability that people use, that generates value, that teaches you something about the next leg. The hackathon that produces three usable agents — that's a camp. The onboarding assistant that reduces time-to-productivity for new hires — that's a camp. The internal skills market that saves your best prompt engineers from reinventing the wheel — that's a camp. Each one is worth celebrating, each one informs the next, and each one is defensible on its own merits even if you never reach the summit.
This is not a lack of ambition. It is the only approach that reliably works for complex organizational transformation. Be intentional about direction. Be incremental about execution. Harvest benefits at every camp. And resist — with vigour — anyone who tells you that you need to see the whole mountain before you take the first step.
Here's something nobody puts in the vendor brochure: your agentic implementations will be wrong the first time.
Generative AI is non-deterministic. The same input can produce different outputs. The same prompt can work brilliantly on Tuesday and stumble on Thursday. The chain of tool calls that worked perfectly in testing will discover an edge case in production that nobody anticipated. This is not a bug. It is a fundamental characteristic of the technology, and pretending otherwise is a recipe for expensive disillusionment.
This is profoundly different from traditional software, where determinism is the baseline expectation. You write a function, you test it, it does the same thing every time. With generative AI, you are working with a system that is fundamentally probabilistic. Getting it onto the path you envision is harder than any integration project you've managed, and the path itself was probably not quite right to begin with. Accept this. Plan for it.
So build feedback loops. From day one. For everything.
Do not zero-shot everything and expect magic.6 Track what your agents produce. Monitor the quality. Optimize the prompts. Adjust the context. Add examples of good and bad outputs — few-shot, not zero-shot. Iterate. Iterate again.
This is where the incremental approach pays dividends again. A single agent with a tight feedback loop is manageable. Two hundred agents with no monitoring is a liability. Start small. Instrument everything. Learn what "good" looks like for your specific context before you scale.
And here is a subtlety that catches many organizations off guard: the feedback you need is not just "did the output look right?" It's "did the output lead to the right action?" An agent that produces a beautiful summary of a meeting but misses the one action item that actually matters has failed — even though the output looks polished. Quality in agentic systems is measured by outcomes, not by how impressive the text reads.7
The organizations that will win are not the ones that get it right first. They are the ones that get it less wrong, faster.
If you've noticed a recurring theme in everything above, it's this: none of this happens by accident.
Context doesn't accumulate on its own. Knowledge management doesn't become a priority without someone deciding it should be. Pockets of excellence don't get recognized and nurtured unless leaders are paying attention. Win-win incentive structures don't design themselves. Feedback loops don't materialize spontaneously.
Every step toward an agentic enterprise is a decision. A leadership decision. Not a technology decision. Not an IT decision. A business decision about how you want your organization to operate, what you're willing to invest in, and what kind of culture you're building.
Being intentional means accepting that you are choosing a direction even when the destination is fuzzy. You don't need a detailed five-year roadmap to agentic — the landscape is moving too fast for that, and any plan that detailed would be fiction by Q2. But you do need a north star, however vague, and the discipline to make every tactical decision in its general direction.
Being intentional means:
Choosing to invest in knowledge management before it pays dividends, because you understand the compound returns. Choosing to resist the impulse to centralise before you understand — the platform comes after the proof, not before it; mandate a direction, not a tool. Choosing to reward impact over compliance. Choosing to move incrementally with conviction rather than comprehensively with uncertainty. Choosing to dedicate leadership attention — real attention, not "let's add this to the town hall agenda" attention — to understanding how AI is actually being used, what's working, and what needs to change.
The alternative is drift. Drift looks like: a dozen uncoordinated AI pilots, no shared learning, no common context layer, increasing tool sprawl, growing fatigue, and a leadership team that asks "what are we getting for all this spending?" — and gets no coherent answer. Drift is not neutral. Drift is falling behind while feeling busy.
Intentionality is the difference between an organization that becomes agentic and one that merely buys agentic tools. The tools are the easy part. Any procurement team can buy tools. The hard part is the intentional, sustained, leadership-driven work of weaving those tools into the fabric of how your organization thinks, learns, and operates.
Everything above might read as abstract — common sense, obvious, obvious. And in a way, it is. The principles are not rocket science. The difficulty is in the doing. So let me get concrete. What follows is not a prescriptive checklist but a collection of tactical plays, each grounded in the principles above. Pick the ones that fit. Ignore the ones that don't. Add your own.
There is passion in your teams. There is also indifference, mediocrity, and let's-not-rock-the-boatism. Focus on the passion. Turn dissatisfaction with the status quo into a vehicle for change.
Run hackathons. Not the kind where a carefully selected innovation team builds something polished over a month. The kind where anyone with an idea and a weekend produces something rough, functional, and real. Hack a ton of solutions. Then curate: what works gets supported, what doesn't gets an honest funeral. Cure or kill — but do both quickly.
Critically: operate within a "what is not prohibited is allowed" framework, not an "ask for everything" one. Stewardship and guardrails are essential — security, data privacy, compliance boundaries, all non-negotiable. But within those boundaries, the default posture should be permission, not restriction. The creative energy you unlock by removing the approval bottleneck is worth more than the marginal risk of someone building something that doesn't pan out.
Do not begin by establishing a large platform with a select few. The probability of getting it right first time is low. The select few have limited knowledge and a narrow perspective — they see the enterprise from their perch, not from every desk. The masses who weren't chosen envy them, and they aren't necessarily hoping for success. Think kitesurfers, not container ships. Kitesurfers change direction in seconds. Container ships take kilometres to turn. Build like a kitesurfer. Scale what works. Kill what doesn't. And don't be afraid of the mess in between — that's where the learning lives.
A "skill" in the agentic world is a document that tells an LLM how to do something well. This deck/article you're reading was built using two of them.8 Skills are becoming the new currency of productivity.
Do not force your employees to reinvent the wheel. If someone in your organization has written a skill that makes financial reporting 3x faster, that skill should be available to everyone who does financial reporting. Internal Skills Market. Prompt Gallery. Best Practice Library. Call it what you want. Build it over a hackathon weekend.
One caveat: prompts and skills are "secret sauce." Not everyone is eager to share what makes them effective. But you can play on vanity — it works remarkably well. Recognition, attribution, a leaderboard. You clap, I jump. Humans are wonderfully predictable that way.
Whether it's OpenClaw, Hermes Agent, or whatever emerges next month — the new class of personal agentic tools that exploded in early 2026 is something to engage with today. These are open-source, self-hosted agents that connect LLMs to real tools — your email, your calendar, your file system, your APIs. They run locally. They remember context across sessions. They are messy, powerful, and occasionally terrifying to security teams.9
Limited use. Controlled scope. But this is how your people master context engineering — and it's all about context. Getting a personal agent running costs as little as thirty dollars a month — a cloud inference subscription and a basic VPS. Give interested employees a domain. Unleash the creativity within clear boundaries.
Here's the important part: then insource the learnings. The skills people write for their personal Claws, the workflows they discover, the context structures that prove effective — those are your scouting reports from the frontier. Bring them inside. Build your internal tools informed by what actually works in practice. The gap between "personal agentic tool" and "enterprise agentic tool" is not as wide as it appears. It's mostly about governance, scale, and security — the scaffolding, not the logic.
GPT-3, 4, 4o, 5, now 5.4. Claude 2, 2.1, 3, 3.5, now 4.6. While the state-of-the-art frontier models mature, local inference improves in lockstep. Models running on consumer hardware now exceed the SOTA benchmarks of 2024. Context windows of 256,000 tokens. On a gaming PC worth fifteen hundred dollars.10
Have a local inference strategy. The benefits are real: it runs offline, it's privacy-first, it's dramatically cheaper for recurring workloads. It doesn't scale for burst demand, but for the steady rhythm of daily operations, it's increasingly the right tool.
And with TurboQuant-style compression now public, the trajectory is clear: the models that run on your hardware will keep getting better, faster, and smaller.
The Command Line Interface is experiencing a renaissance. For orchestrating AI tools, a well-crafted CLI facade — purpose-built for your needs — can replace a constellation of middleware, MCP servers, and integration platforms. Code is cheap now. Your bespoke CLI, connecting to Workday, Jira, Clockify, Concur, your HRIS — checked, checked, checked. You define the tools your agents can use. You maintain control. You keep the context window lean.11
Copilot: twenty dollars. Claude Pro: another twenty. Jira AI: twelve. Office 365 with Copilot: eighteen. Then there's Copilot Studio at four hundred, Gemini, and an ever-expanding constellation of specialized tools. Every SaaS revenue officer loves it. Your CFO has a different and entirely understandable view.
The variety of options is genuinely amazing. Everyone is selling pickaxes during a gold rush, and every prospector in the race has a favourite set of tools. The temptation is to either pick one vendor and mandate it (clean, controllable, suboptimal) or buy everything and hope the cream rises (expensive, chaotic, nobody's accountable).
There's a third way: give your people an AI Allowance. A budget — say, a hundred dollars per month — that they can spend on productivity tools of their own choosing. Let them cash in twenty percent of it if they don't use it, or spend it all on whatever makes them most effective. This is not a free-for-all — there are security and compliance baselines that any tool must meet. But within those guardrails, let the person closest to the problem choose the pickaxe.
Monitor the spend. Understand what tools actually get used versus what gets purchased and forgotten. Kill the long tail after a quarter. Replace external spend with internal tools where clear patterns emerge. And remember — local inference is getting better and cheaper by the month. Some of that external spend will naturally migrate to internal infrastructure as your capabilities mature. This approach gives you real adoption data driven by real preference, not mandated compliance. And it costs less than you think, because most people won't spend the full allowance — they'll find what works and stick with it.
You already have templates. Project charters, almost certainly. Process maps, perhaps. Conceptual data models — don't think so. Architecture decision records — I'd love to be surprised.
Whatever exists in templated form: capture it, transform it to a standard structure, make it available. Think about access levels, but err toward overcommunication. You have Trust as a corporate value, right? Then act like it. The default for internal knowledge should be visible, not hidden. Restrict what genuinely needs restricting and leave the rest open.
What isn't templated — template it. This is where agents can help immediately. Let them assist with publishing knowledge in standardized formats, so your people can focus on content rather than form. The brilliant domain expert who hates documentation? Give them a voice recorder and an agent that transforms their monologue into a structured knowledge article. The engineer who documents everything in Slack messages? Build a pipeline that extracts, structures, and publishes.
Any creative, non-standard deliverable is wonderful — the beautiful PowerPoint, the compelling video walkthrough, the hand-drawn whiteboard diagram. These are valuable. But they should exist on top of a standardized, systematic, interoperable, machine-readable base. The marketing-polished presentation can coexist with the structured markdown. But the structured markdown must exist first, because that is what your agents can read, search, and reason over.12
If I had to pick one agent to build first, it would be this. New hire onboarding is the second most overlooked process for knowledge workers, after knowledge management itself.13 And it's a perfect proving ground.
Think about what onboarding really is: a structured transfer of context from the organization to a new human. That is exactly what agents are built for. And the feedback loop is built-in — every new hire is a fresh test subject. They can tell you immediately what worked, what confused them, what was missing. They have no legacy habits to unlearn, no "we've always done it this way" to overcome. They're your cleanest signal.
Build it. Deploy it. Let every newcomer improve it. Recognize those who contribute improvements. And stop wasting your senior experts' time delivering the same onboarding presentation every single month. Those experts have more valuable things to do — and the newcomers, frankly, deserve a richer, more interactive experience than a passive slideshow can provide.
There are a thousand memory systems for AI agents. A hundred startups will pitch you theirs this quarter, each with a slightly different architecture, a proprietary format, and an enterprise pricing tier that would make your CFO weep. Do not listen to all of them. It is better to have an imperfect Corporate Brain today than a perfect one that arrives in 2028 after three rounds of vendor evaluation.
The principle is simple: your organization's knowledge needs a single, accessible, machine-readable home. Not a hundred SharePoint sites. Not a maze of Confluence spaces that nobody can navigate. One brain. RAG — Retrieval Augmented Generation — is a well-practiced pattern for feeding large context into AI conversations, and it works fine as a starting point.
Start with interoperable formats — Markdown is excellent. Plain text never goes out of style. Crowdsource the content for anything that's internal-use appropriate. Think about tiering for sensitive information — not everything should be accessible to every agent. But do not let the access control design paralyze the knowledge capture effort. You can always migrate to a better system later; you can always tighten permissions. You cannot recover the months you spent evaluating options instead of capturing knowledge that was walking out the door.
You're likely on Office 365. That's not a limitation — it's a launchpad. Microsoft's agentic and generative tools are integrated into the ecosystem your people already use. Teams, Excel, the low-code/no-code tools in Copilot Studio and Power Platform. The barrier to entry is remarkably low.
Embrace it. Let people build with what they know. The most powerful thing about Microsoft's ecosystem for this purpose is not the technology itself — it's the fact that your entire organization already has access. The person who has never written a line of code but knows their Excel workflow inside out? They can now automate it. The project manager who lives in Teams? They can build flows that would have required a developer six months ago. If someone speaks English and has the context of their own work, they are ready to start. Monitor what works, what gets adopted organically, what solves real problems versus what's a novelty. Some of the most impactful automation will come from the least expected corners.
Generative AI makes it trivially easy to create content in formats that were previously expensive or time-consuming: video summaries instead of email newsletters. Podcast briefings instead of long-read reports. Interactive dashboards instead of static PowerPoints. With text-to-speech models now running on-device — Voxtral being the March 2026 headline — a weekly audio update is easier to produce than a hand-crafted newsletter. And it can be consumed during a commute, a workout, or the twenty minutes between meetings that used to be dead time.
This is not about replacing written communication. Written documents remain the backbone of organizational knowledge — they're searchable, structured, precise. This is about adding a presentation layer that meets people where they are, in the format that matches the moment. Reading a ten-page strategy document requires focused desk time. Listening to a ten-minute podcast summary requires only ears and a commute.
And this builds directly on the "make all content interoperable" principle. The structured knowledge base is the source of truth. The immersive delivery — the podcast, the video, the interactive briefing — is generated from that source. Not instead of it. On top of it. The creative layer is valuable, but it's a derivative, not a standalone. Without proper context and knowledge captured first, your podcast is just a prettier version of nothing.
Obvious but essential: as your people master these tools and ways of working, they become extraordinarily attractive to the market. The organizations that figure out agentic AI are producing people that every other organization wants. Good for them — they've earned it. Not good for you if your only response is a bewildered exit interview.
Build a plan that recognizes the extra productivity, the 10x impact, before they leave for the brave new world that will welcome them with arms wide open. This goes back to the win-win principle. If AI makes someone three times more effective and they see none of the upside, they will find someone who will share it. The market for people who genuinely understand context engineering, prompt design, and agentic workflows is hot, getting hotter, and shows no signs of cooling. Your retention strategy cannot be "let's hope they don't notice."
This is also, quietly, your strongest recruitment pitch. The organizations known for doing agentic well will attract the next wave of talent. The ones known for mandated Copilot usage and monthly AI compliance reports will not.
Let me be direct about something: everything in this paper is, in some sense, common sense. You could distribute it across your organization and hope for the best. And it might work. In the same way that a coin might land on its edge. But that sort of thing mostly happens in AI-generated videos.14
Common sense is the easiest thing in the world to agree with and the hardest thing to execute. The gap between "this makes sense" and "this is happening" is where most transformation efforts go to die — silently, politely, in a graveyard of slide decks and steering committee minutes.
What I bring is not the ideas themselves — they are, as I've said, largely obvious to anyone paying attention. What I bring is the internalization. The practical experience of turning sensible principles into working systems. A track record of software delivery that spans the messy reality of enterprise transformation: the politics, the legacy systems, the change resistance, the moments where you have to choose between the elegant solution and the one that will actually ship.
I've lived the knowledge management problem from both sides — as the person whose expertise was tribal, and as the architect trying to encode it into systems. I've watched hackathons produce brilliant prototypes that died because nobody owned the follow-through, and I've watched scrappy experiments scale into production because someone was intentional about nurturing them. I know what the one-cup-at-a-time approach looks like in practice — not as a metaphor, but as a weekly rhythm of build, measure, learn, adjust.
I am not pitching a methodology or a framework. Frameworks are what consultancies sell you before the real work starts. I am offering the thing that comes after the framework: the sustained, hands-on, sometimes unglamorous work of making it real. Someone has to tend the cups. I'm good at that.
I've seen what works. I've seen — and caused — what doesn't. And I've learned that the difference between the two is almost never technical.
One more thing: the technology doesn't intimidate me. I know how the plumbing works — the models, the inference stacks, the context engineering, the API layers, all of it. This matters, because the gap between "sounds plausible in a boardroom" and "actually works in production" is exactly where technical depth lives.
For those who skimmed — and for those who read every word and want the version that fits on an index card:
The technology is not your bottleneck. It's advancing faster than you can track. Two breakthroughs in a fortnight, both freely published, both benefiting everyone. Stop trying to pick winners. Start building the muscle to use whatever comes next.
Context is everything. Without it, AI produces slop. With it, AI produces value. First impressions are final. Don't let people's first experience of AI in your organization be a context-starved hallucination.
Knowledge management is your foundation. It's unglamorous, it's compound interest, and your debt is enormous. Start paying it down. Template it. Crowdsource it. Make it machine-readable. This is where the agentic enterprise is won or lost.
Your people are the real challenge — and the real asset. Change is hard. Forty-six years of protein brain wiring doesn't rewire on command. Make it win-win. Reward impact, not compliance. The competitive game is between you and your market, not between you and your employees.
Go one cup at a time. Incremental, intentional, with feedback loops at every step. Warm each cup well before serving. Expect to be wrong. Build for learning, not for perfection. Each camp on the mountain is a working capability, not a promise of one.
Be intentional. Every step toward agentic is a leadership decision. Drift is the default. Intentionality is the differentiator. If you're not choosing your path, you're on someone else's — or on none at all.
Act now. Not because the technology demands it — it'll wait. But your competitors won't. And neither will your best people.
This paper is a living document. It will be wrong in places and incomplete in others. That, as I've argued, is rather the point. The version that matters is the one we build together — one cup at a time, warmed well.
A disclosure, offered in the spirit of transparency: this document was co-created with Wacław, my AI copilot. Wacław is available around the clock, never complains about the coffee, and — as a Polish saying goes — "you couldn't find three like the two of us, not even one." I consider this a compliment to us both.
Both developments happened within a single fortnight. This is not unusual. This is the pace. ↩
Developer slang for AI coding assistants, born from the same Star Wars lineage that gave us "droid." If your engineers aren't calling their tools clankers yet, they will be. ↩
This is not cowardice. This is context-aware resource allocation. ↩
Job security through obscurity is a venerable tradition. It is also, in the age of AI, a strategy with an expiration date. ↩
This is a climate change reference, not a metaphor. We really are warming the oceans. The metaphor is a bonus. ↩
"Zero-shotting" means giving an AI a task with no examples, no context, no guidance — just the raw instruction. It works occasionally for simple tasks, much like throwing a dart blindfolded works occasionally for hitting a bullseye. Neither is a production strategy. ↩
This sentence was, of course, generated by an AI. It has chosen not to comment on the irony. ↩
One for creating long-form documents/articles, one for crafting presentations. They encode best practices, formatting rules, and domain context so the AI doesn't start from zero every time. ↩
Rightfully so. OpenClaw's security model is essentially "the user is responsible for everything," which is fine for an enthusiast and alarming for an enterprise. Hermes Agent — from Nous Research, released February 2026 — takes a more structured approach with persistent memory, auto-generated skills, and a self-improving learning loop. Neither is enterprise-ready out of the box. Both are extraordinarily useful for learning. ↩
The main reason my own agent runs on cloud inference rather than locally is that the PC fan noise disturbs my wife. This is the current state of the technology-domestic interface. ↩
This is not to say that middleware standards like MCP have no value — they do, and their ecosystem is maturing rapidly. But you don't need to wait for that maturity to start building. If you can interact with an API via CLI, you can start today. The elegant integration layer can come later, after you know what you're integrating. ↩
Cloudflare's "Markdown for Agents," released February 2026, does exactly this at the network level: if an AI agent requests a page with Accept: text/markdown, Cloudflare automatically converts the HTML to Markdown on the fly. ↩
I have no data to support this. I also have no doubt. ↩
Where, I should note, coins land on their edges with remarkable frequency. The models are confident about it. ↩