"Agent" was the word of the season this summer. The tech press filled up with it: ad platforms rebuilt around agents, open networks for them to talk to each other, media shops testing agents that buy on their own.
If you run an agency or sit in a law firm marketing department, someone above you has already forwarded you a think-piece about it and asked what your plan is.
For a lot of firms, and nearly every law firm, agents are a year out, probably more. No law firm reading this is handing a live client matter to an autonomous agent in 2026, and it should not. The risk tolerance is not there, and the infrastructure is not either.
But the question underneath the agent question does not wait a year. Whatever tool you eventually adopt, one thing decides whether it produces work you can use or work you delete: the context layer underneath it. And almost nobody is funding that.
Think of your AI capability as a building. The tool everyone is excited about is the penthouse. Top floor, best view, the thing you point at when guests come over.
The context layer is the foundation. Three levels down, in the dark. It does not photograph well, it will not be in the announcement, and it is the only part that keeps its value while everything above it gets replaced.
Watch what happens when a team buys the tool first. Someone signs up for the platform everyone is talking about. The demo is clean.
Week one, a few people generate a few things, and the output comes back sounding like nobody at the firm. For the agency, it is a byline that reads like a press release when the client's whole brand is dry and understated. For the law firm, it is a client alert that is sure of itself about a practice area in a way that would make the practice group lead wince.
So the person who owns the work rewrites it by hand, the way they always did, and files the tool under overrated. By week three they have stopped opening it.
A badly configured tool kills a team's enthusiasm faster than almost anything. The tool was never the problem. It just did not know a single thing about who it was writing for, so it wrote for no one.
That is the tell that you started on the wrong floor. The AI did not fail. It succeeded at producing a draft with no fingerprints on it, which is worse.
So what is this layer, concretely.
For an agency, it is the house voice for each client written down as rules a machine can follow: the words the client bans, the way their CEO sounds in a byline, the difference between how you pitch a tier-one reporter and how you write an award entry.
It is the messaging framework and the boilerplate you have already fought over and approved. It is the library of pitches that earned a reply. It is the case studies your team rebuilds from scratch every time a new-business request comes in.
For a law firm marketing and BD team, it is the attorney bio in the firm's voice, the practice group description the partners have already argued over and signed off on, and the representative-matters list that business development pulls into every pitch and RFP response.
It is the firm's style rules. How the firm name appears, which words are house style and which are banned, the citation format nobody wants to fix by hand. It is the tone that separates a client alert written for a general counsel from the same update written for a founder.
And it is the governance. What can be said in public and what cannot, the confidentiality lines, the no-guarantee-of-outcome rules a junior associate would not think to flag but the firm cannot cross.
The best description I have heard for all of this is a second brain. The firm's judgment and voice, captured once, available to everyone.
The distinction here is simple. A one-off prompt gives you one good output from one clever person on one good day. Context gives you a consistent output from everyone.
It also builds inside the work itself. An AI that holds onto the matter and the client history means the fortieth draft is smarter than the first instead of starting cold every time.
None of that is a product you buy. All of it is work you do. Which is exactly why it gets skipped.
The reason the order matters is that the two layers age in opposite directions.
Whatever tool wins the next eighteen months will be obsolete in the eighteen after that. The agent everyone is demoing today, the platform your peers are piloting, whatever your firm eventually approves, all of it gets leapfrogged or repriced by the next thing.
That is not a knock on any of it. It is the whole trajectory of the category. The tool is a rental, and the rent resets constantly.
The context layer does the opposite. Every persona you write down and every workflow you capture is yours. All of it grows more useful as it accumulates.
Swap the tool underneath it and the context still works.
This is exactly why the year-long wait is not a reason to sit still. If you spend this year writing down your voice and the work you are proud of, you are ready the day the tool you trust finally arrives. If you wait for the tool first, you do all of that work anyway, under deadline and badly.
In agency after agency and firm after firm, I run into the inverse of this. Three smart people each build their own clever setup. None of it is written down, none of it is shared, and when one of them leaves, the capability walks out the door with them.
That is not an AI capability. That is three private hobbies that happen to use AI. Individual productivity that never turned into institutional capability, which is the only kind that shows up on a balance sheet.
I run a firm that helps organizations build exactly this layer, so weigh what I say accordingly. But I did not land on this because it is what we sell. We built the business around it after watching every stalled rollout turn out to be missing the same thing.
If I could put one sequence on the wall of every agency and every marketing team staring down an AI mandate, it would be this.
Context first: the voice, the personas, the approved positioning and the best prior work, written down and shared. Workflows second, built on that context so they produce drafts that already sound like you.
The tools third, agents included, once there is a foundation solid enough to hold them. Measurement running underneath the whole time, so you can separate the workflow that earns its keep from the one that only looked good in the demo.
Do it in that order and the costly, exciting layer comes last, standing on something that can hold it.
Do it backwards, tool first on nothing, and you get the thing teams pay for all the time: an impressive demo and a team that now believes AI does not work for them. That belief is the real cost, and it runs far higher than the license.
The newest tool can wait a year. The layer underneath it is what to start on now.
Build down before you build up.