The Stenographic Mediator
The thing AI makes possible that Agile could only promise: a silent intermediary that sits between every expert in a software organization, translates each one's dialect, shows each only the concerns that touch them, and lets a unit of work clear every reviewer in parallel instead of through a chain of meetings. Plus the part no one else is building, a pipeline that turns juniors into seniors on purpose.
The first piece in this series argued that AI does not solve software's oldest problem, the way intent degrades as it passes from stakeholder to manager to developer. It inherits it, and executes the inherited misunderstanding at machine speed. It ended on a promise: that the most valuable use of AI in a software organization is not as a code generator, but as the translator the industry has always needed, never managed to reliably hire, and rarely valued when it found one. This piece names that pattern and defines it. I have spent a decade building company architectures, and this is the thing I kept wishing existed.
What Agile was actually for
Strip away the ceremony, the stand-ups and sprints and story points, and Agile had one real goal: to create a shared language between the business and the people building the software. The reason it is needed is simple and rarely said plainly. Domain experts cannot reliably convey technical constraints, and technical experts rarely hold deep domain knowledge. Both are true at once. The rare person who is fluent in both has always been extraordinarily valuable and almost impossible to hire, and the industry has never properly recognized or paid for that role.
Agile brought the right people closer together. It could not make them fluent in each other's language. So it delivered the next best thing it could: meetings. A lot of meetings.
You are already doing this, just not on purpose
Here is what makes this urgent rather than speculative. Your organization is already running its delivery process through AI, right now, whether you sanctioned it or not. Your product managers are drafting half their emails and requirements in ChatGPT. Your developers are prompting models to write code, and the ones who do not even think of themselves as AI users are Googling the same questions and reading the AI-generated answer at the top of the page as fact. Then those developers ingest the poorly structured, now AI-shaped communications the product managers produced, and feed them into their own tools. The telephone game the first piece described is still running, except now there is an unaccountable model sitting at half the nodes.
None of it is observable, structured, or governed. You cannot tell what a model generated from what a person wrote, or which decisions were an AI quietly filling a gap with a confident assumption. The context lives in a hundred private chat windows that vanish when the tab closes. The fidelity loss the first piece described is now happening faster, in more places, with no one holding the thread. That makes it wasteful, the same work redone in a dozen tools and never reused; dangerous, ungoverned generation flowing into code and contracts with no review trail; and costly, every private session a metered call nobody is accounting for.
The Stenographic Mediator does not introduce AI into your delivery process. AI is already there. The mediator is the difference between that happening in the open, observable and routed and preserved, and it happening in the dark.
That is the opportunity hiding in the warning. You cannot put this back, and you do not need to. You can accept that your organization is already doing it and then build structure around it, structure that makes the same AI use observable, routes it where it actually helps, and keeps it from being wasteful or destructive. Done deliberately, the thing already happening starts working in your favor instead of against you, and without taking on major new risk.
The pattern: a stenographer who is also bilingual
Picture a court stenographer. They record a room with perfect fidelity, they can read any part of it back on request, and they do it in whatever form the room needs, all while never participating in the conversation. They serve it.
Now put an AI in that seat, between every party who has a stake in a piece of software: the stakeholders who want it, and the experts who have to shape and approve it, in design, engineering, infrastructure, operations, legal, governance, finance. Three things become possible that never were before. I call the pattern the Stenographic Mediator.
It translates. Each party speaks its own professional dialect. The mediator converts between them. When engineering asks "do we need federation," the stakeholder hears "should businesses outside our company be able to log in with their own email and password," and answers a question they can actually answer.
It filters. Each party sees only the concerns that touch them, rendered in their own language. Legal does not wade through deployment topology; the backend team does not read the marketing brief. That is the part that kills information overload, which is its own silent tax on every large project.
It parallelizes. A unit of work, a single story, goes to every concerned party at once and gathers their requirements, questions, costs, and sign-off, instead of crawling through a sequence of meetings where everyone waits their turn.
Why this is the thing Agile promised
Put the concerned parties on one axis and the stories on the other. In a normal organization the work moves down that grid sequentially and synchronously. Everyone has to be in the meeting. The meetings happen in order, while other people wait in planning. Ideas that made the thing work evaporate because no one took notes or wrote a ticket that survived the week. Priorities shift mid-stream and leave half-built work on the floor. By the time intent reaches implementation it barely resembles what the stakeholder described in the first meeting.
The Stenographic Mediator makes the same work move across that grid in parallel and asynchronously, and nothing is lost, because the mediator holds all the context and can read any part of it back to anyone at any time in their own language. A legal requirement on a page becomes a placement decision for design, becomes a prototype for the front end, becomes a question back to legal in plain English, all without a meeting and without a dropped thread. When every party has signed off, the story enters the backlog fully specified. The stakeholder only ever sees the two things they actually care about: time and cost.
And the parties on that grid are not just reviewers. This is the blackboard pattern: each discipline contributes its own invention to the shared story, in its own field, in parallel. Legal does not approve a compliance section someone else wrote, it writes the compliance approach. Infrastructure designs the topology. Security shapes the threat model. The story becomes the sum of real expert work from every layer, assembled without a single meeting.
This is collaboration without the ceremony. It is the promise Agile made and the overhead it never managed to shed.
The honest part: what is new here, and what is borrowed
It matters to be straight about this, both because it is true and because it is the only way the idea earns the right to a name.
The translation-and-routing machinery underneath this pattern is a recombination of ideas software has known for decades. The mediator that centralizes communication so parties never talk directly. The canonical model that lets you translate everything through one shared form instead of building a separate translator for every pair of dialects. The anti-corruption layer that keeps one discipline's model from quietly corrupting another's at the boundary. The scatter-gather and blackboard ideas that let many specialists enrich a shared artifact in parallel. An integration architect will recognize all of it. I am not claiming a new primitive, and anyone who tells you their AI workflow is a brand-new computer-science discovery is usually selling something.
That is the point worth holding onto, not a caveat to rush past. These patterns were always sound. What kept them out of the actual delivery process was never the ideas, it was the overhead of running them by hand. A mediator by hand is a meeting. A canonical model by hand is documentation nobody updates. Scatter-gather across ten disciplines by hand is ten calendars and a month of latency. The patterns were proven and impractical at the same time. The one thing that has genuinely changed is that an AI can carry the overhead these patterns always demanded, which is what moves them off the whiteboard and into a workflow you can actually run.
What is genuinely new is two things. The first is applying all of that to translate between professions rather than between data formats, with each human getting their own rendered view of one shared truth. The second is the part that matters most to me, and it is below.
Making seniors on purpose
This is the direct extension of the first piece in this series, AI development is a communication problem, which asked where the senior engineers of 2040 will come from once the industry has cut the junior rung that produces them. The Stenographic Mediator answers that as a byproduct of how it works.
Because this system breaks work down to a far finer grain than a normal backlog, you can sort that work by difficulty. That unlocks something the industry is currently destroying.
You can hand graduated, deliberately unassisted work to junior developers, matched to where they actually are, and track it. And you can deliberately route the consequence-bearing projects, the painful migration, the architecture decision that has to live with its own past choices, to the people who are ready for them. Those are the moments the first piece in this series called anagnorisis, the recognitions a developer can only earn by living through the absence of good design, the reason interfaces and modularity and dependency injection stop being academic and start being obvious. They cannot be taught with a lecture. The consequence has to be experienced. It works the way a ballet company grows a lead dancer: it makes sure the dancer gets the demanding, career-defining roles, a lead in The Nutcracker, exactly when they are ready for them, not before and not too late. This does the same for engineers, inside the delivery process itself.
The answer is here. The system that delivers the software also grows the people who will one day be senior enough to catch what the AI gets confidently wrong. That is the piece with no prior art, and the piece I care about most.
There is a mirror image to it. The junior pipeline grows new expertise from the bottom, one consequence-bearing project at a time. The same fine-grained structure lets the people already at the top of a field write the path forward from above. Because every discipline is present on every piece of work, you can feed the frontier of each one into ordinary design, the newest ideas from the people leading it, clearly marked as experimental and free to fail. One side of that coin raises up the next generation of experts; the other lets today's experts raise the standard everyone builds to. The delivery system becomes a knowledge engine in both directions, not just a faster way to ship what is already known.
Composable expertise
One more benefit worth naming. The concerned parties do not all have to work for the same company. A legal-review specialist, a governance firm, a UX studio can each plug into a story behind the same translation interface, respond in their own dialect, and never interact with anyone outside their lane. Expertise becomes something you compose per project rather than something you hire permanently. A stakeholder could, in principle, put a feature into the system and get back the only two numbers they wanted, time and cost, after every discipline between the idea and the running software has weighed in.
The infrastructure is arriving, and it is not all in the cloud
There is an obvious objection to all of this. A Stenographic Mediator is a heavy AI system. It runs a translating agent for every concerned party, continuously, in parallel, and its core activity is constant handoff between those agents as one discipline's dialect becomes another's. That is an enormous amount of inference. If every one of those translations is a metered call to a cloud API, the mediator becomes the exact rent-extraction layer it was supposed to remove, a new tollbooth between a business and its own software.
This is not the first time computing has run this loop, and the loop has a name. The mainframe era kept the machine in a guarded room and gave everyone else a dumb terminal, a screen and keyboard with no power of their own, wired back to the center. It looked permanent. Then minicomputers, then workstations, then the personal computer, and the capability that once filled a room moved onto a desk and then into a pocket. The phone you are probably holding runs several thousand times faster than the Cray-2, the most powerful supercomputer on earth in 1985. Computer scientists have a name for the underlying pattern: Bell's Law of computer classes, Gordon Bell's 1972 observation that roughly every decade a new, lower-cost class of machine forms and then grows powerful enough to take over the tier above it (Bell, Communications of the ACM, 2008). The industry has already lived the full swing once, from mainframe to client-server PC to cloud and now back out toward the edge. Centralization keeps turning out to be the first phase of a capability, not its final shape. Cloud AI is the mainframe phase. The terminal is your browser. The workstation is arriving now.
Two shifts arriving right now make it feasible to run the bulk of this on infrastructure an organization owns, rather than renting all of it from a hyperscaler.
The first is local hardware. The 2026 wave of AI laptops and mini-PCs, NVIDIA's RTX Spark, AMD's Strix Halo, and the DGX Spark workstation, can run capable models locally, on a desk, for the price of the hardware instead of a per-token bill. They are best at exactly the kind of model the mediator leans on: sparse Mixture-of-Experts models, which read only a small slice of their weights per word and so run fast on consumer machines. I wrote about why that is eroding the cloud's lock-in in Arming both sides. The mediator's per-party translation is stable, high-volume, latency-tolerant work, which is precisely the kind that can come in-house. These translating agents, the stenographers in the metaphor, can live on the machine each person already uses, part of their own device, rather than a service rented and reached for across the internet.
The second is an efficient wire. A mediator is the textbook case of agents talking to agents, the pattern I argued is where AI's hidden waste compounds in The envelope tax. Every dialect translation is one model emitting tokens that another model immediately consumes; shipping that back and forth as text wrapped in web packaging taxes every hop for a reader that does not exist, because both ends are machines. A token-native protocol like Codec strips that packaging, so the mediator's constant internal chatter costs a fraction of the naive cloud version and can even ride edge and low-bandwidth links. Codec also offloads work from the center: the small dictionary lookups, conversions, and safety pre-checks a cloud used to perform can happen on the edge device that is already there, so the same machine running the agent absorbs that work instead of paying a data center to do it.
None of this means the whole system leaves the cloud, and it would be dishonest to claim it does. A central coordinator, the part that holds the shared canonical state and routes work between the parties, is very likely to stay in the cloud, and the heaviest frontier reasoning may stay there too. What comes off the cloud is the expensive, repetitive bulk: the constant per-party translation, the working dataset, and much of the downstream work, all of which can run on hardware an organization owns. The cloud's role shrinks from metering every word to coordinating a few. That is the difference that matters, between a tollbooth on every exchange and a lightweight switchboard.
Put together, the Stenographic Mediator is not a bet that cloud AI gets cheaper, and that is deliberate, because the evidence runs the other way. The economics laid out in Five Siphons point in one direction, and it is up. The electricity the cloud runs on is rising, the chips are scarce and capacity-constrained, and the hundreds of billions already committed in capex have to be recouped through pricing or written off as losses. A metered cloud layer sitting in the middle of your delivery process is therefore a cost that grows with use and with time, on infrastructure whose own costs are being forced upward. So the expensive bulk of the mediator is buildable on hardware a company can own, behind a protocol that does not tax its every internal exchange, with only a lean coordinator left in the cloud. That is the precondition for it to be a productivity tool rather than one more metered layer between a business and its own software. The infrastructure to move that bulk out of someone else's cloud is arriving on the same timeline as the pattern itself.
What this is, in one line
The Stenographic Mediator automates the actual goal of Agile, shared understanding across disciplines, by stripping out the ceremony and the fidelity loss, and it turns the same fine-grained view of the work into a pipeline that grows the next generation of senior engineers. I am defining the mechanics formally, and building it, at Quasarke. This piece plants the flag. The implementation is its own story, and a later one.
It is worth saying plainly that this is not a someday idea. It is already happening, just without structure. AI is loose in your organization right now, in a hundred private windows, creating more chaos than coherence and burning money at every desk. The question was never whether to let it in. It is whether to give it a shape. Handled deliberately, the same force making a mess today becomes the thing Agile promised and never delivered, fluency across every discipline without the meetings, and a pipeline that keeps producing the seniors who hold it all honest. That is a sustainable way to work, not just a faster way to lose the thread.
The first piece in this series opened with a joke: three people walk into a bar, one with industry knowledge, one with technical knowledge, and the third holding the entirety of human knowledge but still somehow needing the other two to get anything done. The Stenographic Mediator is how you finally get all three working at once, in parallel, across every dialect in the building, without anyone having to sit through the meeting.
William Dunn is the founder of Quasarke, a software development consultancy, and has spent a decade building company architectures. This is the third piece in a series on communicating with AI as a practitioner.