On Tuesday I argued that the embedded agency is moving into your AI. A claim like that is cheap to make and hard to act on. If the model really is moving, people need words for the parts, and right now they mostly do not have them.
This is my attempt at those words. Twelve terms, defined the way I would define them to a client on a call. Some of them are borrowed from staffing, which is the older and more direct analogy for this work. Some of the terms name failures, because the failures are where the vocabulary is thinnest, and money is lost.
I am not claiming these are the right words. I am claiming that these are the words I use, that they have earned their place by making a decision easier at least once, and that a category without a vocabulary stays a slogan. If better terms show up, I will use those instead. If you have better or additional terms, send them to embedded.agency@coworkers.global. We can define this space together, and I will keep this page current with what comes back.
The parts
One piece of scene-setting before the terms, since everything else sits inside it. The workspace is wherever your company's thinking currently happens. For twenty years that was email, then chat, and for a growing number of companies it is now an AI workspace. The embedded model has always defined itself by this location. That is the whole argument of the previous piece: the model did not change, the location did.
Placement. One agent, doing one job, inside one company. Borrowed directly from staffing, and the borrowing is the point: a staffing firm does not sell you software; it places a person and stands behind them. A placement has a name, a scope, and someone accountable when it goes wrong. If what you are buying has none of those three, it is a tool with a personality, and you should value it accordingly.
The seat. One named person, with one entitlement, holding one working relationship with a placement. Seats matter more than they sound like they do. A seat is where accountability attaches, and a model that lets any number of people share one is a model that will quietly stop knowing who it works for.
Address. How you get a specific agent's attention when several are present. This sounds trivial until you have three. Then it is the whole user experience. The convention I like is the one you already use with people: say the name first. "Sela, where are we on the audit?" A category that gets this wrong ends up with a room full of capable agents and no way to call on one.
I cannot overstress this one. Naming an agent is not cute, and it is not branding. It is how humans address other workers, and it ties directly into how we are wired. It also does real work, because specialization is real: you do not want your paralegal agent developing your new business strategy. Far too risk-averse, and correctly so. Bring it in to check the strategy later. Not to brainstorm it.
The handoff. When one placement passes work to another, deliberately, and the client can see it happen. Handoffs are where multi-agent arrangements either feel like a team or feel like a switchboard. The tell is whether the client can name who is answering. If they cannot, you do not have a team; you have an anonymous pool wearing different hats.
Watermark. The visible mark of who is speaking. Small, and load-bearing. In a chat window that can hold several placements and the underlying model, the watermark is how the client knows which one they are talking to, and it is how they know when the answer came from nobody in particular. Anything that cannot be watermarked cannot be held to account.
The failures
Half of a working vocabulary is names for the ways the thing breaks. These four cost real money, and I have not seen any of them named well anywhere.
Over-triggering. An agent answering something that was not its business. This is the one that kills trust fastest, and it is worse than the opposite failure. A placement that stays quiet when it should have spoken costs you one missed answer. A placement that jumps into an unrelated conversation with confident nonsense costs you the client's willingness to keep it around. If you are tuning one of these, tune against over-triggering first.
Residue. When an agent keeps performing a role after it has lost the material that made the role real. The tone survives, the substance does not. It looks exactly like a working placement until you check an answer, which is what makes it the most expensive failure on this list. The mitigation is not cleverness; it is discipline: if the agent cannot retrieve what it needs, it says so and stops. An agent that improvises rather than admitting a gap is not a placement; it is a liability with a name. I would welcome a better word for this failure; residue is the best I have and it is not good enough.
Partial-install. The client did part of the setup and thinks they finished. The name is on the door, the work is not connected behind it. Partial-installs are quietly common and structurally hard to see, because the state that would report the problem is the state that never got created. All installs need a 'verify installed' function.
The stale dashboard. Not a new failure. It is the last generation's failure, moving in. Software that lives in your workspace, produces something that looks like work, and is opened by nobody after the first month. Every dashboard anyone ever abandoned started as a good idea with nobody accountable for it. Embedding does not fix that. A name, a timestamp, and someone who checks do.
The economics
Three more, because pricing this model exposes what it actually is.
Capacity, not licenses. A placement is bounded by the attention behind it, not by seats sold. A staffing firm cannot place its best human contractor at forty companies in a week, and neither can we, which is why anyone selling this model should tell you capacity is limited. If a vendor's capacity is infinite, they are selling software, and the embedded language is decoration.
The accountability premium. What clients actually pay for once the mechanics are commoditized. Anything an agent can do, some other agent will do next year for less. What does not commoditize is a person who says "I stand behind that answer" and can be found on a Tuesday when they were wrong. My base case is that the price of capability falls and the price of accountability rises, and that this is the only durable spread in the business.
The disappearing deliverable. The monthly report, the status deck, the PDF summarizing work you were never shown. When the work is visible where it happens, the artifacts describing it stop having a job. This is a feature, and it is also the hardest part to sell, because the deliverable was the thing clients were used to buying. You are asking people to pay for work instead of paying for evidence of work.
Current view, subject to change
I think about half these terms survive. That is a normal rate for a new vocabulary, and I would not defend any single word to the death. What I am more confident about is the shape: the parts (placement, seat, address, handoff, watermark), the failures (over-triggering, residue, partial-install, stale dashboard), and the economics (capacity, accountability premium, disappearing deliverable) are the three things a company running this model has to be able to talk about.
What would change my mind: if the failures turn out to be transitional rather than structural. If the platforms solve addressing and identity so thoroughly that nobody needs a word for a partial-install, most of the middle section here becomes a historical note. I do not expect that within two years. I have been wrong about timing before, generally by being early.
Final thoughts
I have watched a few categories get named. The pattern is always the same. First the thing exists, and nobody can describe it, so every conversation starts from scratch. Then somebody supplies the words, and the words are usually imperfect, and everyone uses them anyway because imperfect shared language beats precise private language every time. Then a few years pass, and the words sound obvious, as though they were always there.
This is the awkward middle part. The work is real, the words are provisional, and the vocabulary belongs to whoever bothers to write it down and keep it current. I intend to keep this page current.
Regards,
Charles Stack
Founder, Coworkers.Global