Tuesday afternoon, four o’clock. An engineer has been three hours into a problem and emerged with what he went in for: a migration plan, a trace through an unfamiliar service, a diagnosis of why the obvious approach was about to break. The work is “done.”
He opens Slack to explain it to his teammate and forty-five minutes later, he’s still typing. The message will be three paragraphs at most, any longer and his teammate won’t read it, any shorter and he can’t capture his wins. He tweaks the first paragraph endlessly and has rewritten the second four times. The temptation to paste the session into an LLM and let it summarize lurks constantly, but he knows in his bones that the summary would be missing something. So he hand-rolls something because this is “as good as it gets.”
Throughout this process, a nagging voice asks when he’s going to get back to the real work. And the voice has a point: he isn’t building anything right now; he’s just “rearticulating” what he already worked on. His message will take at most five minutes for his teammate to read, assuming he’s paying attention, and summarizing it should be “easy” because “the hard part is already done.” The voice is wrong in a way that takes years for most to articulate then refute.
The truth is that the value can’t be captured in five minutes. “The work itself” is only part of what needs to be done, and that took three hours to produce by rejecting approaches, noticing a small surprise that turned out to be the key, and divining the exact way in which the obvious answer is a trap. The “rearticulation” work is actually the work of building a tiny replica of his mental model inside his teammate’s head, using only words. It is at least as hard as the three hours that came before, usually harder.
He can feel that it’s work because he’s exhausted when it’s done. His attention has been spent, but the voice in the back of his head, which he has not yet fully trained out, remains unsure.
The adage says you can go fast alone or far together. But it doesn’t say what “going far together” costs each person in mental model construction. The cost is not about collaboration per se, which everyone values. We’d all be lost without colleagues whose competencies we lack (e.g., the PM who knows the customer personally, the designer who can see the system as a stranger does, or the senior engineer whose terse design reviews save months of wasted work). Each of them sees a facet of the gem the others cannot. None of us can be all of them.
Our protagonist (and most professional culture) is missing something crucial in his narratives about “work” and “value.” Persisting what we’ve come to know in another mind is not only one of our defining characteristics as human beings, but also one of the most significant endeavors we undertake. It is and always has been extremely expensive, directly dependent on the speaker’s working memory, entwined with wisdom and experience, of elders around a fire.
We’re increasingly impatient: the AI-induced compression of time-preference makes the cost now feel unbearable. Meanwhile, AI’s collapse of execution costs has made this the dominant remaining transaction cost within firms. The naive answer is to move more to agents, but that only exacerbates the problem.
Nothing is wrong with the human mind. The substrate of modern work needs to grow.
Costly Comms
Communication, in the technical sense, is two operations stacked on top of each other. The sender serializes, encodes a thought into a string of symbols. The receiver deserializes, decodes those symbols against his own mental model. If the two mental models share enough structure, the deserialization is cheap. Imagine someone explaining a bug he found in a service he’s been living inside for weeks. He says, “the timeout logic is backwards.” To a colleague who’s also been in that service, staring at the same code, those five words light up the exact machinery the sender used to produce them: the recipient sees the bug immediately; he knows which timeout, which paths trigger it, and why it matters.
But the same message to someone who’s never seen that service forces the recipient to ask: which timeout? What does “backwards” mean in this context? How does it connect to the rest of the system? Every reference in the sender’s message has to be re-grounded against unfamiliar terrain. The recipient is building a mental model from scratch while the sender assumes he already has it. The cost of re-grounding is paid in the working memory of both interlocutors: the recipient’s, trying to hold context that could have been in the message; and the sender’s, trying to fill in the gaps. In interactive exchanges, both pay continuously. Working memory is the most expensive resource in the human brain. It’s where thinking actually happens. When we ask someone to deserialize a message against a model he doesn’t have, we implicitly ask him to spend his most precious resource on this reconstruction, a necessary precursor to, but not itself, understanding.
Many software engineers will recognize this as sharing the shape of “side effects.” Consider a service whose job is simple: route incoming requests to the right backend based on the request path. The routing logic itself is trivial: a few hash lookups, maybe a regex match. Five microseconds of actual work. But the service is also instrumented. It logs every request so operators can understand traffic patterns, debug weird behavior, and audit who accessed what. That log is written to disk, and disk I/O, even to a fast SSD, is measured in milliseconds. Suddenly the service spends ninety-five percent of its time on the side effect (logging) and five percent on the work the service is nominally for. The routing didn’t get slower. The return value is identical. What changed is the cost of communicating that the routing happened. In a well-instrumented system, the side effects often dominate the total cost. A function that takes a microsecond to compute might spend a hundred microseconds writing the log line that says it computed. The work of communicating that it ran is what took most of the time.
Human communication is a side effect of thinking. That is, thoughts are return values of the mind that are recursively fed back for further thinking and action. The cost to produce the thought is whatever it was, but the distinct work of producing the artifact that lets another person have a similar thought is unbounded in the way computational side effects are. It can cost tens or hundreds of times the thinking itself. And, like logging, it does not change what was thought, but what can participate in it.
This cost is not visible to most people because most people are inside their own heads when they’re “doing the work.” From inside, the references resolve cheaply: the thing I was looking at when I had the insight is right there. The decision I rejected three minutes ago is still fresh in my memory. The sender dereferencing each of these pointers pays nothing for their resolution, while the receiver, dereferencing in a different head, not only pays for every pointer but must also reify a facsimile of each referent.
This is why writing a clear summary is so much harder than producing the thinking the summary is about. The thinking happened cheaply because the substrate was already there. The summary has to manufacture the substrate from scratch, in someone else’s head, using only the words that fit in a Slack message. It is also why the cost of communication shows up in our bodies as exhaustion rather than on our calendars as time. We don’t feel tired because we spent three hours “writing.” We feel tired because we spent three hours holding two contexts at once: the problem we’re solving, and the imagined reader who will have to solve it again from our artifacts.
Csikszentmihalyi named the state in which working memory is fully deployed against the problem — fully, with no spare capacity bleeding off to maintain context for an imagined later reader — flow. Flow is what happens when we stop splitting our attention. It’s the felt signal of minimized friction, the state in which the barrier between intention and action dissolves. We think something and it happens; the work moves at the speed of thought because there’s no translation step in between.
The serialization tax destroys flow specifically because it taxes the same resource flow requires: working memory, attention, and the capacity to hold a complex model without losing pieces. Most of us can’t be in flow and be writing our own stenographer’s transcript at the same time. The moment we start asking ourselves, “how do I explain this to someone else,” we’ve split our attention, and part of our working memory is now devoted to translation rather than thinking. The substrate that requires the transcript is the same substrate that breaks flow. And once flow is broken, the exhaustion we feel isn’t from the work itself, but of maintaining two minds concurrently.
Pierre Menard
There is a Borges story that makes this concrete in a way that no engineer’s diagram can. It is called “Pierre Menard, Author of the Quixote,” and a friend of mine (Joe Warren) introduced me to it when I was trying to explain why writing the same sentence in two different organizational contexts produced two entirely different outcomes.
The premise of the story is that a twentieth-century French symbolist named Pierre Menard sets himself the task of composing a few chapters of Cervantes’ Don Quixote. Not copying the chapters, but composing them: word-for-word identical to the original, but produced from inside Menard’s own life, his own century, his own situation. The narrator then treats Menard’s text as a genuinely new work and analyzes it as such. The same sentences, in Cervantes’ mouth, are seventeenth-century commonplaces. In Menard’s mouth, they are dazzling philosophical provocations from a man who has read Nietzsche and William James. Same characters on the page, yet different works.
The story is funny, and the joke does real work. The joke is that an artifact, a sequence of characters, does not contain its meaning. The meaning is reconstructed by the reader, dereferencing the artifact against everything the reader brings to the page. Two readers with different contexts can read the same characters and produce different works in their minds. There is no fact of the matter about what the artifact “really” says, because the artifact, by itself, doesn’t say anything. The reader is what says.
This is “the curse of knowledge” in geometric form. The curse of knowledge is the cognitive bias in which a person who knows something cannot accurately imagine the perspective of someone who doesn’t. Most of the time, we explain this bias in language that makes it sound mild (e.g., “they forgot what it was like not to know”), but the Menard frame shows what’s actually going on. The sender of an artifact has one Pierre Menard in his head: the version of the text that reads against the sender’s actual context. The receiver has a different Pierre Menard. The two Menards can hold the same characters and produce entirely different works, and neither party is in a position to notice the gap because each party is reading their own Menard and assuming, charitably, that the other is reading the same one.
In one head, the sentence is a witty deflection; in another, a sincere commitment. Same words, different meanings.
Summaries as Fun House Mirrors
The standard collaboration move, when you finish a piece of work and need someone else to act on it, is to summarize. You compress the work into a paragraph, a page, or a fifteen-minute readout, and you hand it off. The recipient reads the summary and forms a view of what was done. That view is built from two ingredients: your compressed artifact, and the recipient’s own model of the world. You can see your contribution. An experienced writer can guess at his readers’ models, but a guess isn’t “access.”
A summary is a fun house mirror, but with an important asymmetry: a real fun house mirror is honest about what it is. You see the curve in the glass; you know your reflection is stretched, compressed, and warped, so you discount what you see accordingly. But a summary doesn’t show its curvature: it presents itself as a flat mirror, so the recipient reads it and assumes he is seeing the work undeformed.
The deformations are not accidental but inherent. A summary compresses by length, which forces choices about what survives. It tends toward narrative coherence, which smooths the loose ends the work actually had. It asserts what the work suspected, dropping the confidence calibration. It presents the conclusion without the chain, so the reader gets the verdict but not the reasoning. Trying to document the reasoning doesn’t fix this: summarizing reasoning is itself a metacognitive act, expensive for the sender to perform and still incomplete for the recipient. Each deformation is a curve in the glass the recipient cannot see.
And that’s only half of the distortion. The recipient brings his own model to the reading, and the Pierre Menard problem stacks on top of the mirror. The deformed reflection is read against a model that deforms it further. By the time the recipient finishes the summary, the work he holds in his head is twice removed from the work that produced it, and neither he nor you can see how far either deformation has carried.
Most of the time, the gap doesn’t surface at handoff. The recipient nods. The conversation feels aligned. He repeats some of the sentences back at you, and the sentences sound right, because the sentences are the part you can both see. The model he built, the thing that will actually drive his next decision, is invisible from outside his head. You will not learn that it diverged from yours until a meeting six weeks later, when he makes a call you would not have made, and you realize, watching him argue for it, that he has been reasoning from a different premise the whole time. He didn’t lie and he didn’t misread: he read the reflection in the only mirror he had, against the only model he had, and the work he constructed in his head was a different work.
This is “the curse of knowledge” operationalized at firm scale. The cost of summarizing is not the time spent writing or interrogating the summary, but that of the misaligned downstream decisions, paid every time the deformations matter. The worst feature of this cost is that you cannot detect it without paying it. There is no diagnostic that runs at the moment of handoff and tells you whether the recipient’s model matches yours. The only diagnostic is the consequence, which arrives weeks or months later, by which point you have long since moved on from the work and lost the context that would let you properly course-correct.
Most would conclude, “we should have aligned earlier,” but that phrasing is wrong. “Aligning earlier” would have meant more summary: longer, more careful, or spoken aloud instead of written. None of those change what a summary is. In other words, the mirror would still be deformed and the recipient would still not see the curve. The depth of the alignment is bounded by the fidelity of the medium, and the medium is the problem.
Editorial Distance: Social & Temporal
Two adjacent framings are helpful when considering the challenges of writing for oneself and others over time and space.
Social: The Stenographer’s View
The stenographer I have in mind isn’t a courtroom typist. Picture instead the academic department secretary of an earlier era — late 1960s through the 1980s — highly educated, often more so than the men whose memos she types, ever-present in their meetings and invisible to them in equal measure. She is physically inside the room. She hears every word. But she is not in the conversation.
The very position that makes her invisible to the men in the room is what makes her useful. She sits close enough to hear everything and far enough outside their shared context to know what they cannot see themselves omitting. The participants speak tersely, eliding what they take for granted; she, just outside that shared context, fills in what they assumed when she writes up the notes for circulation. Her record includes exposition the participants would never have written, because her reader was not in the room and because she has learned, over years, to write for readers who never were.
When a participant writes a note about the same conversation, he writes from inside it. He assumes his reader was inside too. He omits exactly what the stenographer would have included. “You were there, weren’t you?” is the implicit question in every terse note ever written by a participant.
No, Ted. He wasn’t. That’s the problem.
The note is doing the wrong job. It is a record made by a participant, for participants, handed off to someone who is not a participant. The exposition that would have made it legible for Ted was never on the writer’s list of things to include, because the writer’s mental model of the reader was someone who already shared the room. The participant cannot easily be the stenographer at the same time he is participating — doing so requires modeling an outside-position reader from inside the conversation. This is the same two-minds-concurrently cost from earlier, in a particular shape. What the artifact omits, it omits not because the writer was sloppy, but because from inside, nothing he omitted seemed missing.
Temporal: Sisyphean Proanamnesis
The first framing was about position. The second is about time, and it applies even when the writer and reader are the same person.
By the time your colleague reads what you wrote, or by the time your future self reads what your past self wrote, your context has already moved on. You no longer hold the state of mind that produced the artifact. You’ve moved to a new problem. Your working memory has been cleared and refilled. If the artifact is ambiguous (e.g., if your colleague reads “use approach D instead” and isn’t sure why D was chosen), you cannot retroactively inject the missing context. The version of you that knew the context, that held the three rejected alternatives warm in working memory, no longer exists in a form that can be queried.
A common objection is that the limitation is theoretical and that careful writers overcome it in practice. However, most of us are not in a position to notice how badly we forget, because we can’t introspect our own memory failures across time with any accuracy. The proof is in the reminders: we have admitted, at the level of the alarms we set for ourselves, that we forget routinely. And these are things we explicitly told ourselves to hold onto. If even the loud, deliberate effort to remember fails this reliably, the silent, ambient remembering of context-without-flagging-it stands no chance.
The standard answer to this is “write better forward-looking notes.” But notice what that asks of you. It asks you to know, at the moment of writing, what your future self will have forgotten. It asks you to anticipate the gaps in a context you don’t yet have, on behalf of a reader you don’t yet contain. There is a word for this practice (I had to invent it), proanamnesis, from Greek pro- (πρό, “in advance”) + anamnēsis (ἀνάμνησις, “the act of recollection”). Fore-recollection. Recollecting in advance what hasn’t been forgotten yet. An English sister, if you prefer the portmanteau, is in futrospect — future + retrospect.
The word is a coinage because the practice doesn’t exist. Greek and Latin between them have spent thousands of years naming subtle operations of memory, anticipation, recollection, and writing-for-posterity. There is no word for anticipating what your future self will have forgotten and pre-empting the forgetting from the present. The hole in the lexicon is the trace of an impossibility.
This is the Sisyphean version of “write better notes.” Sisyphus was the mythological king of Corinth who, for his offenses against the gods, was sentenced to roll a boulder up a hill for eternity — the boulder rolling back each time he neared the summit. The labor was not just difficult; it was structurally futile, ordained to fail at every iteration. Writing-better-forward-notes is the same task in cognitive form. Each new project, each new artifact, each renewed resolution to this time, document things well enough that future-me will know what past-me meant, ends with the boulder rolling back. The future-me who reads the note has forgotten exactly what past-me would have needed to write. Past-me could have known what to write only by being future-me, which past-me could not have been.
By attempting this, we become our former selves’ “stenographers in futrospect” — and bad ones because the stenography is a side effect we never agreed to.
Strangers to Ourselves
Pierre Menard addresses the interpersonal aspect of the problem while the Sisyphean view addresses the temporal one, your future self as your colleague’s spatial counterpart. Both are directionally correct, but they understate the full extent of the problem in two ways.
First, the serialization tax is not a binary that activates only at handoff between people. It is a continuum that begins the moment a thought leaves the head that produced it. There is a tax at five seconds, and a tax at five days, and the tax at five weeks is just a larger version of the tax at five seconds. The reader you are writing for, even when you write for yourself, is already a stranger; she just hasn’t gotten very far away yet.
Second, pair-wise communication is the easy case. In practice, almost nothing in an organization is pair-wise: a Slack message goes to a channel, a design doc circulates to a team, a leadership memo lands in every inbox. Each recipient brings his own Pierre Menard, builds his own facsimile, traces his own deformation. Everything established for the pair-wise case applies in every direction at once, and a new failure mode emerges that has no pair-wise analog. The receivers drift from one another, not just from the sender, and none of them is in a position to see it happen. The cost is quadratic in audience size, not linear: every pair of heads in the room is a potential misalignment, including pairs that never involved the sender at all.
Nothing is Certain but Death and Taxes
The two standard responses to the serialization tax both accept the medium as given and respond by adding work upstream of the boundary.
If my notes were better, this would be fine. If I disciplined myself, the way the conscientious people on engineering blogs say I should, into producing rich, well-structured artifacts at every juncture, then handoff (to others, or to my own future self) would resolve cheaply. It would not. The serialization tax is not a craft problem, so it is not solved by working harder on the artifacts you produce. The tax is built into the medium. Better notes raise the floor; they do not change the slope.
“Write better docs” tries to lower the tax by raising the quality of the artifact: every sentence is denser, every paragraph more packed, every diagram annotated, and arguments longer and more careful. The recipients still have to deserialize it against their own models. Although the image is higher fidelity, it comes at the expense of the author’s time as he guesses about the shape of his audience’s varied mental models. Furthermore, the deserialization now takes longer because the artifact contains more material to dereference. The total cost of the channel has increased. In return, the alignment improves marginally and falls off fast because the receivers’ models are the bottleneck, and writing more does not change the receivers’ models. At best it can anticipate approximate gaps.
“Communicate more” tries to lower the tax by amortizing it across additional contacts: daily standups, weekly syncs, async write-ups, or OKR check-ins. Each contact costs a serialization step on the sender’s end and a deserialization step on the receivers’ ends. The hope is that small frequent updates produce alignment that one large update couldn’t. They attempt to capture the limits of human attention and the blast radius of long-running misalignment problems, but at the price of large running costs on both ends and a substantial cumulative tax on flow. Most organizations have found steady blood letting more tolerable to emergency surgery that arises from ad-hoc discovery of misalignment.
There is a structural reason both these responses fail, though. Charlie Munger, Warren Buffett’s longtime partner and the more aphoristic half of Berkshire Hathaway, famously said “Show me the incentives, and I’ll show you the outcome.” If the right behavior (i.e., proactive, rich communication) is more expensive than the wrong one, you should expect, over a long horizon, that the wrong one wins. This is the incentive gradient in action: people are not bad; the gradient is bad. Asking conscientious individuals to fight the gradient is asking them to volunteer for friction in perpetuity, and they will burn out doing it, they will quietly stop doing it, or they will succeed at it for a while and then leave the company, often to the detriment of the firm.
There is a third, more sophisticated response than the first two: build culture, shared vocabulary, jargon. Amazon’s leadership principles, mechanically, actually are a semantic compression algorithm: Two-pizza team, Disagree and Commit, Working Backwards. Each phrase that needs no explanation. When two Amazonians say “is this a two-way door?”, they’re invoking a shared frame with which each was explicitly inculcated. The serialization is cheap because the receiver’s substrate already contains the structure the sender is pointing at. Deserialization is cheap for the same reason. The jargon is the substrate, externalized into vocabulary.
This is what culture actually is, when you strip away the warm-and-fuzzy framing: shared mental models and expectations. Inside jokes, leadership principles, internal acronyms, the way the company talks about itself — all compression. The whole point of culture is to make “2 > 0” carry meaning without having to derive it from first principles every time. It’s worth noting that every sufficiently large organization will eventually produce a culture, mostly accidentally and mostly still with rampant inefficiency. But a culture nonetheless because it’s how communication scales beyond the size where everyone can hold everyone else’s full context in their head. With a strong culture, communication moves faster, decisions are made faster, and alignment is strictly cheaper. The firm can be larger because the coordination cost has been pre-paid through cultural investment.
Increased velocity, however, has broken culture’s ability to ameliorate mental friction. The mechanism worked when communication happened slowly enough that the serialization tax was amortized across days or weeks; decisions came at a pace humans could track, and culture-as-compression could keep up. Even with the best jargon in the world, work is being created faster than even the best compression can compensate for. The limiting factor is no longer how clever your vocabulary is, but the bandwidth of the human brain itself.
This hard biological limit has implications for the future of firm structures and scales and how units within firms will compose productively. In exploring why firms exist at all rather than individuals contracting freely on the open market, in 1937, Coase (who won the Nobel for this work in 1991) answered that markets are not free to use. Every transaction has costs: finding the right counterparty, negotiating terms, verifying that the work was done, and enforcing the deal when it wasn’t. When those costs grow large enough, it becomes cheaper to put people on payroll and coordinate them internally than to keep transacting in the open market. The conclusion is that firms are what arise wherever in-house coordination is cheaper than out-of-house transactions.
The corollary is that the cost structure of coordination determines firm shape. When coordination is cheap (i.e., when people can find each other, communicate easily, and build shared understanding quickly), firms grow. They can afford to be large. And when coordination is expensive (i.e., when moving information between people costs time, effort, and accuracy), firms shrink. They fragment into smaller units that can operate more independently or quietly degrade as their internal transaction costs eat their margins, producing less than they should for the resources they consume.
Knowledge work today has shifted because execution costs have collapsed: agents have amplified individuals into teams, so the cost of doing the work is no longer the binding constraint. Now, costs of coordination dominate and ironically worsen with deeper adoption of AI: more output volume (decisions to align on, artifacts to integrate, etc.) increases the surface area of coordination, and with it, opportunities for misalignment. This cost now defines what firms can do at what scale. The serialization tax of moving thoughts between heads and building shared models across people, paid every day by every conscientious person on every team is the dominant remaining transaction cost in a post-AI firm.
The evolution of the optimal firm size will be determined by affordances of the communication substrate. The way out is not to add more work upstream of the boundary, but to change what the boundary is.
A New Substrate
Instead of translating the work into a different medium for transmission, we should preserve the work as it happened. Increasingly, work is done in close collaboration with an agent such as Claude Code. This presents an opportunity to allow colleagues to resume and review sessions that others have had from any moment, at the live state of the thinking, rather than receive a translation of it in a parallel channel. The conversation about the work should happen at the work, not somewhere else.
This requires a few primitives, each of which replaces something current tools do poorly.
Ambers
The foundational primitive is the amber: a preserved snapshot of an agent session. Every substantive piece of work in Claude Code (or a comparable agent) is a session: the back-and-forth between human and agent, the tool calls, the rejected paths, the eventual conclusion. In current tools, this session evaporates when the window closes. A transcript may survive, but the state (i.e., the live model of the work, the place where the thinking left off) does not.
An amber preserves the state. It is not a transcript of what was said; it is a snapshot of where the thinking was when it was preserved, which means it can be warmed — opened again as a working session that resumes from exactly that point. The agent that warms an amber doesn’t read about the prior work, but it picks up from exactly where it stopped.
Every other primitive in the substrate anchors to ambers. An amber is essentially a session-wrapper onto which a pin is attached, from which acorns are generated, around which burls form, and through which the grain runs.
Acorns
When an agent investigates something on your behalf, it uses tools: WebSearch, file reads, API calls, database queries, Grafana queries. The PM researching how Tableau represents a particular view triggers a dozen WebSearches and reads through API docs. The engineer planning a refactor across nine repos triggers file reads and greps and pulls in Slack threads for code context. The teaching assistant preparing materials for a class searches GitHub for example repos and pulls research papers. In each case, the work is based on the “tool outputs” the agent collects on the way to producing whatever the human asked for.
In current tools, those outputs evaporate and ground truth is lost. None of the API pages the agent read, the search results it weighed, or its snippets persist. Three weeks later, when the PM revisits the same question or when a colleague comes asking about the research, the work has to be redone from scratch. The investigation is invisible to anyone but the human who watched it happen (if they were even watching).
With acorns, as agents use tools, their outputs are captured automatically and structured into a queryable record. No marking or human action is required: acorns form around the exhaust from work the agent was doing anyway. By recording a snapshot of the data used to make a decision, acorns make the agent’s conclusions auditable. When Claude produces a finding (e.g., “the ICP skews toward mid-market with team sizes between 50 and 200”), the human can drill into the evidence that supports it: which surveys, interviews, etc.. He can double-check and dive into the agent’s sources to extend the research further. Agent-facilitated research goes from being black-box outputs to traceable claims with provenance. Findings are the raw material for artifacts that humans share (e.g., strategy docs, briefing memos, proposals, or root-cause analyses).
Acorns encapsulate a pipeline from evidence to findings to artifacts. Artifacts are the visible work; findings are structured intermediate output; and evidence is primary sources. Today’s artifacts seldom have any provenance; with acorns, the artifact ships with everything it was built on.
Pins
Pins are the small things you mark along the way. You’re tracing a bug through three services, deep in the work, fully in flow. You’re looking at service A’s request handling, and you notice something tangential: service B is mis-handling the timezone on a field nobody is using yet. It’s not the bug you came to fix. It’s not blocking anything. It’s just something you see, something that will matter in three months when someone actually uses that field.
That observation today presents a conundrum because our current tools all trade flow for persistence.
Option One: ignore it. You’re in flow, so you keep going. The observation flickers through your working memory and disappears. Three months later, someone hits the bug and spends a day debugging something you already knew about.
Option Two: break context and write a ticket. You open Jira, you describe the situation in a blank form, you add it to the backlog. You’ve just paid a context-switch tax to preserve something that wasn’t your main work. You finally return to your original work, but are forced to re-ramp up. Flow: shattered.
Option Three: Slack it. You send a message to the team channel pointing at what you found. It lands in a stream of hundreds of other messages. Someone might see it, might not. If they do, they have only a glimmer of what you’re referring to because for you to have clarified would have required the same flow-shattering engagement as the ticket from Option Two. For example, which service you were looking at, what state it was in, or why you noticed this particular thing? The message gets lost in noise, most readers unwilling to invest the activation energy to unpack an underspecified issue.
Option Four: branch now. Open a new tab, a new chat, a new doc, and start a parallel thread in Claude Code on the tangential thing. This is probably the least disruptive to flow and most helpful to your memory when you return, but you’ve also committed to managing another open context. By the end of the week you’ll have five such contexts open; by the end of the month, fifteen. You’re no longer working but shepherding threads.
Pins offer a fifth option: you mark a turn in your current amber as it happens with a memo for why it’s noteworthy. This action is small enough not to break flow yet structured enough to find again. Because the marker is embedded in an amber, only a light note is necessary to remind others or “future you” of what to pursue further. Hours or days later, when you have capacity for the tangent, you warm the amber at the pin and branch from exactly the context you marked. The branch is deferred, not foreclosed. Your context-management problem doesn’t accumulate, because nothing is open; markers wait until you’re ready for them.
Current tools force temporal commitment at the moment of noticing: you must lose the opportunity, or keep it by breaking flow, broadcasting, or branching now. Pins decouple “noticing” from “acting.” The cheap operation (marking) happens at noticing-time; the expensive operation (branching and engaging) happens at branch-time, whenever that is.
Burls
A collaborator can join at any amber. When he does, and a real argument breaks out, the closest form in the current tools that argument takes is a sidebar: a side thread, branched off the main conversation, where multiple people deliberate in parallel to the work. What’s actually happening in that moment is much richer than “sidebar” would suggest, so we call it a burl.
In a real tree, a burl forms when the tree encounters damage, disease, or pressure: something in the world that demands a response. Rather than growing normally, the tree responds by thickening at that exact point, concentrating resources into a denser, more resilient structure. The grain at a burl is compressed, intricate, harder than the surrounding wood. It’s the tree’s physical record of an encounter with nature, transformed into strength.
In an organization, a burl forms the same way: not from routine work, but from an encounter with something real and hard. The thinking densifies around a genuine conflict between multiple defensible positions, (e.g., a new market entrant or new cost constraints). Multiple perspectives collide and compress. The judgment concentrated at that point is harder-won, more intricate, and carries the organization’s actual values and reasoning in a way smooth consensus can’t. That contact with reality is where insight emerges and is what forms a company’s grain.
The difference between a sidebar in current tools and a burl is locality to context. A Slack sidebar floats free of the work that prompted it. To start one, one has to leave the work, describe the situation in a fresh context, wait for people to load that context, debate, reach a conclusion, and then go back to the work and try to remember where one left off. A burl is the same conversation, but anchored at the amber where it began. Collaborators can get up to speed more quickly and engage more intelligently because the burl is at an amber point that can be branched or interrogated. With each contributor researching independently from a shared understanding, semantic alignment is near-automatic. The conversation and each investigation that informed it are adjacent to the exact context that spawned it.
Differentiated Heavy Lifting
Woodworkers prize burls and seek them out because the grain is more beautiful, more intricate, and more valuable where the stress was. It’s even called “character” from the Greek for “engraving,” an immutable mark.
Incidentally, burls essentially are an organization’s alpha: in that exact character, especially because agentic work is an undifferentiated commodity. Smooth execution and frictionless flows can be done by anyone now. Even agents can do that. Operational or execution complexity doesn’t matter anymore. Instead, resolving difficulty is what produces value. Stress builds character.
The prized wines of France don’t come from regions with abundant water and perfect conditions: they come from places where the vine has to struggle, where scarcity forces concentration. Napa wines, by contrast, grow in ease, and ease shows. The same is true of everything worth valuing: wood, art, organizations. The character that commands a premium, that resists commodification, that actually matters is forged in encounter with the world, not in smooth passage through it.
Grains
Burls resolve into decisions. In current tools, they’d be written down somewhere like a Jira comment, an ADR, or a Slack message. The familiar move is to capture the conclusion, but the conclusion alone is not the unit you want preserved. We call what you actually want, the conclusion plus the chain that produced it, a grain.
In a real tree, the grain is the record of growth over time. Each ring is a year. Each pattern of compression carries the signature of what the tree was doing during that period: wide rings in seasons of plenty, narrow ones in years of stress, dense whorls where the tree responded to damage. You can read a tree’s life by looking at its cross-section, its wood containing the record.
In your organization, grains work the same way. Each is one mark in the larger record, attached to the chain of reasoning that produced it: the amber that surfaced the question, the burl where deliberation began and alternatives were argued. Current tools often produce, “we decided something but I can’t remember why” because the decision and the reasoning live separately, if preserved at all. A grain is the same decision, but it stays attached to its chain. Essentially, attachment is what makes it useful later.
The difference between a decision and a grain is provenance: a decision in Jira is a sentence while a grain is a sentence with a path back through every premise that made it the right sentence. A future reader, human or agent, doesn’t have to reconstruct or speculate; instead, they simply read.
The stack as substrate
The components above comprise a stack, which is the new substrate of agent-native work: ambers as preserved sessions, acorns as exhaust generated from them, pins as moments anchored within them, burls as the deliberations that form at those moments, and grains as the artifacts those deliberations produce. Every node is structured, which means every node is legible to a machine. The past’s Slack threads, design docs, PR diffs, and standup notes are far less so. An agent given an amber, a pin, a burl, and a grain has ground truth; an agent given the unstructured equivalent has only material to confabulate from. Format is a precondition of navigability and utility.
Structure also buys you connection. Currently, when you try to understand a decision, you follow a Slack thread that references a doc that mentions a PR comment that cites a conversation nobody wrote down. The search space explodes; working memory exhausts holding context that should have been in the artifact. With grains connected as a graph (i.e., each decision linked to the burl that produced it, the burl to the amber that anchored it, and forward to the consequences), the search space collapses. The reader, human or agent, traverses a path of explicit nodes instead of guessing at hidden ones. In unstructured search, efforts to find related information resort to using adjacent time window filters or the author’s name as correlates for relatedness. With a new substrate, orthogonal connections central to the mechanism of work but distinct from the artifact itself are automatically constructed and surface what has always existed but is invisible and unappreciated until it’s needed.
Each grain is a mark in the larger “grain:” one decision contributing to the cumulative pattern that defines a company. That pattern is its heartwood, and until now it has lived in places that don’t preserve it well: in Slack threads that age out, in design docs that drift from reality, in the tacit knowledge of people who eventually leave. The substrate is the first medium that makes a company’s heartwood legible: to a new hire, who can read not just what was decided but why; to an agent, which can check whether the reasoning still holds; and to your future self, who will struggle to reconstruct the nuance of your earlier intentions from his or her own cognition and memory. The heartwood is the company’s IP. The substrate is what lets you see it, preserve it, and build on it.
Your heartwood is your competitive moat precisely because it contains your specific struggles, judgments under pressure, and particular taste crystallized by difficulty. No agent can replicate that. It exists only because you encountered something hard and responded with human judgment.
Chesterton, Checkable
The deepest reason this matters is something Chesterton noticed a hundred years ago and that we struggle to apply today.
Chesterton’s fence is the proposition that you should not remove a fence whose purpose you do not understand, because the fence was probably placed for a reason, and the reason may still hold. Imagine you’re walking through a field and you encounter a fence. It’s old, in the way, and you don’t see why it’s there. The conservative reading of Chesterton says: don’t remove the fence. That it takes energy to produce anything and that it exists suggests that the fence was placed for a reason: maybe it was keeping livestock in, marking a property boundary, or protecting something on the other side. If we don’t know anymore, we should leave it. Be slow to undo; keep the fence until you know why it was.
The trouble with the standard reading is that, in real organizations, you almost never know what the fence was for. A decision gets made (e.g., we won’t allow commits directly to main; we require code review; we’ll only hire engineers with five years of experience; or, we’ll run migrations synchronously before deployments). The decision is made for a reason: someone got burned, saw a problem, and wanted to prevent a failure mode they’d experienced. They built a fence. It has worked. Months and years pass. The person who made the decision moves on, maybe even leaves the company. The original problem, the failure mode that motivated the fence, either got solved by something else, or it’s so old that nobody remembers it, or it was never written down clearly in the first place. And then someone new comes in and sees the fence and asks: why do we do this? The answer should be in a ticket, or a design doc, or a Slack thread. But the ticket got closed, the doc was never written, and the Slack thread has aged out of retention, so the fence’s purpose is unrecoverable.
Organizations are left with a binary: either leave every fence in place forever, accumulating decisions that calcify into culture, ossifying the organization, making it slower and less adaptive; or remove fences carelessly and pay later when the original concern resurfaces, when the failure mode that prompted the fence happens again and you realize, too late, why it was there. Most oscillate between the two and call it judgment, but they’re actually just oscillating between paralysis and disaster because the information that would let them choose wisely has disappeared.
A grain at an amber point makes future fences Chesterton-safe.
The artifact is the exit of a full chain. You can see the session that produced the question. You can see the amber where the deliberation began. You can see the burl where the alternatives were argued. You can see the grain, the decision that resolved the deliberation, and the chain of reasoning that produced it. The fence is no longer an opaque relic. The fence ships with its own justification.
When a later actor, human or agent, wants to know whether the fence still has purpose, they don’t have to guess. They simply read the chain. They can ask: do the premises still hold? If yes, the fence stays. If no, the fence comes down, safely, because the reasoning is preserved alongside the decision and the obsolescence is now explicitly falsifiable.
This inverts the Chestertonian default. Instead of “cautiously constrain when you’re sure forever”, teams can constrain freely, because every constraint ships with its own obsolescence check. The mechanism is compositional. Organizations can carry more decisions without paralysis, because every decision is removable on principle. The fences are not eternal. They have a clear life cycle, anchored to whether their reasoning still applies.
But there is something more active happening too. As a company grows and changes, it has to prune. It has to decide which parts of its history still serve its future. The grain of the wood tells you where the stress was, where the growth happened, which decisions were formed under which pressures. When you have visibility into that grain, pruning becomes intentional. A tree left unpruned allocates resources indiscriminately: every branch that can reach light grows toward it, and every sucker (those shoots that sprout from the base or along the trunk, trying to redirect the tree’s energy) draws resources away from the main structure. Trees’ strategies are generally simple: most just grow where growth is possible. Pruning stops that by forcing the tree to concentrate its growth in chosen directions. A company that can read its own grain can prune with the same precision: removing the patterns and decisions that don’t serve its direction, starving the resource-stealing branches, and forcing the organization to grow the way its leadership has decided rather than the way inertia pulls. The grain is the company’s memory of itself. Pruning is how it actively shapes that memory into the future it wants.
This phenomenon isn’t new to companies, but that human beings’ unique ability to navigate ambiguity allowed us to paper over the fidelity deficiencies endemic to our systems. As impressive as modern agents are, compared to human beings, they struggle with ambiguity and reconstructing reasoning from incomplete information as a human would (the reason for this is fundamental and fascinating, but the topic of a separate essay. Suffice it to say that this is not merely a model or context size issue but an intractable misalignment in the spaces that human beings and LLMs each navigate). Despite these differences, we can enable agents to perform what has been a uniquely human endeavor by giving a structured substrate. That is, an agent, given the current codebase and an old grain, can analyze both: the grain says we decided X because of premises P1, P2, P3 and the agent can check whether P1, P2, P3 still describe the current state of the world. If they do, the fence stays. If they don’t, the agent flags it for review. Chesterton-safety becomes machine-checkable. The audit that previously required a senior engineer with two years of context to perform becomes a thing an agent can run continuously, against every decision in the heartwood, at the speed and scale that agents operate.
This reflection and continual dynamic refinement of the organization is not well-served by unstructured exhaust. An agent given a Slack thread and a PR diff cannot do this audit, because the premises aren’t in the thread and the decision isn’t traceable to the reasoning. The audit requires the substrate.
Going Fast, Together
The introduction of this new substrate doesn’t replace collaboration or reduce communication. Instead, it lets us concentrate our humanity where it matters most.
The engineer doesn’t stop doing the work of letting others participate with him. What changes is that “the work” is no longer two distinct, expensive steps. The mental model built across three hours is in the substrate: every amber passed through, every burl formed, and every grain earned. Every artifact is now augmented with the work, not merely a translation of it. The concern of “getting back to the real work” is assuaged. The serialization step discounted as not-quite-work is now absorbed into the work.
The same shift happens for everyone he works with, in their direction. The PM who knows the customer no longer spends forty-five minutes writing a Slack message that turns three weeks of conversations into a paragraph the engineer can act on. Her sessions with the customer are in the substrate; the engineer enters them at the amber where the misalignment surfaced and sees what she saw. The designer who sees the system “as a stranger does” no longer compresses her observation into a Figma comment that has to survive being read out of context. Her pacing-through-the-flow is in the substrate; the moments where she stopped are addressable. The senior engineer’s terse design review is no longer the survivor of a thirty-minute internal deliberation about how to phrase it; her reasoning lives at the burl where the junior engineer was uncertain, and the junior engineer reads not the verdict but the chain. Each facet of the gem stays attached to the hand that polished it. The grateful receiver no longer has to reconstruct what the giver paid to give, and the giver no longer has to pay quite so much.
The reader (colleague, agent, or his future self at nine in the morning next Tuesday) gets what the writer had, dereferenced cheaply against a shared substrate, because the substrate is finally shared. The flow each person has inside their work is the flow they have when someone joins.
“Going fast” and “going far” stop being a tradeoff. The adage was right about what the world looked like in the medium it described. With the right tools, each person can go far at the speed they used to be able to go fast, and the going-together costs less than the sum of their solo work.
Sapwood is the substrate.
This essay names the collaboration tax and the substrate that lifts it. The work continues. The tools that come after the ones we use today will not be refinements of Slack or improvements on Docs; they will replace them the way writing once replaced orality as the central medium of knowledge work, and that precedent — the historical evidence that medium shifts reshape cognition rather than merely accelerate it — deserves its own treatment. And underneath everything here, older than agents and sharpened by them, is the question of whether the work humans do is the same kind of work agents can do at all, or whether the two operate on manifolds distinct enough that the right question is not whether agents transform like humans but whether operations on one manifold land as intelligible side effects on the other.