What are today's AI agents trying to do in an agency?
(note: I let my AI write the first draft of this because of the use of numbers.)
The agents have landed… Strange creatures that seem to want to mimic us, and maybe we should worry about them replacing us. Will it happen? And when? A recent announcement illustrates how AI can replace the parts of your agency that can be written down...and many of them now have been.
The GitHub project is called agency-agents. As of right now it has 282 agents that cover most of an agency org chart. All the things you would expect, especially tech and platform work, but also the likes of account strategist, brand guardian, content creator, and to nobody's surprise ever, project manager.
Every agent file follows a similar structural template that includes defining who it is, the rules it must follow, its workflow, its deliverables, how it should be measured.
If you follow my work, you know I've been writing for well over a year on the potential displacement of agency work/workers by AI. Well, it's here.
Dissecting the Alien Before the Invasion
I literally treated this as if we were exhuming some sort of corpse from a crashed alien spaceship. These AI agents are not an agency yet, but they are the debris of the idea that an agency could be built with agents. So we did an autopsy, Claude and me.
My initial idea was to scan it and see what's most "AI-able" by the frequency and depth of the agents across the various functions. I've been telling agency leaders that agencies should be mapping their different services and functions inside the organization in a way that helps them discover how replaceable they are.
This is basically a reverse engineering version of that, but I also have to believe that this collection of agents also reflects what people have been able to think of as the ways to automate things. When you dig through them, you will find that there's a lot of very specific stuff but also a lot of general stuff lightly brushed over.
I went really far with this project, and did a very detailed analysis of how the scope of the agents relates to actual agency job descriptions, activity levels, and applied it to overall organization, models of various size. I will report more on that later, and in fact ended up developing a methodology for helping agencies assess the potential impact of AI on their organizational shape.
The First Impressions
The whole thing runs 537,066 words. My book, UNMANAGED, ran about 86,000, so that's six novels of agency practice. At a normal reading pace you'd be at it for 36 hours. At least.
The shape of the roster hints at who built it. Engineering is 85 agents. Games and immersive add 27, GIS another 13. Account and client services gets 4. PR gets 1. It's an agency the way its developers see it.
And a lot of it is not horrible. "Never truncate a bar chart's baseline." "Hook in 3 seconds." The Account Strategist's rule: "Never pitch expansion to a customer who is not yet successful with what they already own." Plenty of agencies learn that one the expensive way.
In the terms I've been using, I can see both assembly plus the judgment that rides along with it: how the work gets made, and whether it was made well. In my recent article, I put that at about 70% of what an agency does. In one way, this shows a good deal about what is known about how agencies operate inside of the language model corpus. The distinction between assembly and judgment is a powerful one, because it predicts that assembly work will be the easiest to codify into agentic behavior.
What can AI agents actually do in an agency?
The files are thickest where there's a rulebook to copy. Healthcare Marketing Compliance runs 4,638 words. Legal Billing & Time Tracking runs 3,585. The 22 industry agents are the thickest category in the repo, and what they hold is regulation: mortgage disclosure timelines, ICD-10 coding, trust-account rules. Twenty of the 22 are jobs inside the industry, like loan officer or billing clerk. So basically one industry, and none of them covers how to market in that industry. The Business Strategist covers Healthcare in one line.
The creative core is thin. Content Creator is 335 words, with no rules and no workflow steps. Growth Hacker is 340. All seven paid media agents have no rules and no steps. The thick files are procedure; the thin ones are jobs that are mostly deciding, and nobody could find much to write. I'm going to assume that a lot of these were built using AI, and well, where there is nothing to say, there is nothing to say.
When a file does try to say something about judgment, it comes out one of three ways. A platitude: "Judgment still matters." "Empathy before everything." "Pricing comes after value." A question with the answer left out: "Challenge the assumptions — what does leadership believe that may not be true?" Or nothing at all, like the Content Creator.
The counts back it up. "Judgment" appears 47 times across 25 agents, and 18 of those are in the Search Relevance Engineer, where it's a technical term. "Taste" appears once: "You do not redesign for taste." Tacit knowledge comes up in six lines across four agents, and four of those lines are about fictional characters or training design.
The files also say what they think of deciding. Ten agents warn against gut feel; the Pipeline Analyst is "allergic to 'gut feel' forecasting." Where judgment does come up, it gets handed to a person. The Mobile Release Engineer puts it in six words: "Robots for repetition, people for judgment." The only agents that list their own limits, the 13 GIS agents, route every one of them to another agent. None routes to a person.
So most of the agents are judgment-free. The implied theory is that the language model will supply the judgment on its own. That's a weak bet. The model's native judgment is the average answer, with nothing about any specific client in it.
Some (judgment) assembly required.
The deliverable templates hold 3,489 bracketed blanks. Most are what you'd expect: [Name], [Date], [High/Med/Low]. About 465 ask for something else.
The Account Strategist's QBR prep has a section for who's attending, and under it: "[Who should be there but isn't — and why]." Under industry context: "[Customer's market conditions, competitive pressures, strategic shifts]." The Business Strategist's template asks for "[The one thing that must be true for this to work]," one of 112 blanks in that file. Brand Guardian's brand foundation has 57, starting with "[Why the brand exists]." The Organizational Psychologist asks for "[Founding myths; heroes]."
Those are the questions a good account lead walks into every meeting with. The agent knows to ask them. It has no way to know the answers, because the answers were never written anywhere.
The template even has a place for them. It's called Learning & Memory, and 93 agents open it with the same line: "Remember and build expertise in:". A list of headings follows. The Business Strategist's includes "Decision patterns — how does this leadership team actually make decisions?" Nothing sits under the heading, and nothing in the repo stores what an agent learns. The Senior Project Manager's identity line says "You remember previous projects, common pitfalls, and what works." There's no mechanism by which it could.
Some of what belongs under those headings could be written down and just never was. Some of it can't be. This is what's known as the problem with a specialized class of knowledge that includes tacit, experiential, contextual, and situated knowledge, and many of these are poorly represented linguistically, yet they are critical decision-making methods; these are the areas where humans know, but have difficulty describing. They are the province of those who have what we call expertise beyond book learning.
So judgment shows up in these files four ways: as a blank, as an empty memory heading, as a handoff to a person, and as a rule against intuition. It never shows up as content.
So what are they trying to do?
They're trying to write an agency down. What they've written down so far is a chunk of the codifiable part.
That part is going to keep growing. A lot of agency practice is written down somewhere and just isn't in the repo yet: agency account management, new business and pitching, production and traffic, estimating, media planning, creative and art direction (zero mentions of either), and agency economics like utilization, rate cards and client profitability. Someone will add those files. That's the coming battle, and the line moves up every time a contributor adds one.
To see how close the agents come to real jobs, we pulled 423 agency job postings across 15 roles at junior, mid and senior levels, and rated all 5,061 responsibility lines as either formulaic (follow the procedure) or situated (it depends on this client, this moment). Then we put the agent files through the same test. The agents came out 71% formulaic. The real jobs came out 33%.
Inside the real jobs, the formulaic share drops as you go up: 47% of a junior’s responsibilities are formulaic, 22% of a senior’s, while situated judgment climbs from 9% to 34%. So the agents aren’t describing an agency. They’re describing a junior who never gets promoted: one that is more procedural than even the most junior real job, and missing most of the judgment that real postings ask for at every level.
Objects in This Analysis May Appear Bigger Than They Seem
The judgment blanks in the agent swarm are even larger than they appear. Making judgment a workable method requires a knowledge framework, not a stack of files. I've seen it in action, thanks to the fine work of CEO Bret Starr at The Starr Conspiracy. He built a working system for B2B content generation, and the hard part wasn't how many files it had. It was the knowledge and judgment framework: the structures for representing how the work actually works, then filling it in (contextualizing) for each client's situation.
Bret could build that because he's been doing B2B content for decades and has taught its principles to a lot of people. And he had a running start: I'm pretty sure Bret started on this about four years ago.
Not Today: Five Takeaways
- The 282 agents have written down a chunk of the codifiable work, and that chunk will keep growing. Everything about running an agency that’s documented somewhere will get a file eventually. That’s the coming battle.
- Any given agent’s depth tracks how much procedure exists for the job, not how much value it creates. Compliance and billing get 4,000 words of rules; the creative core gets 300 words and none.
- The agents are mostly judgment-free. Judgment shows up as a blank, an empty memory heading, a handoff to a person, or a rule against intuition, never as content. What’s left is the model’s native judgment, which is the average answer for no client in particular.
- Mostly, they describe a junior who never gets promoted. The agents are 71% formulaic (assembly-focused); real agency jobs are 33%, and even junior roles are only 47%.
- Making judgment work takes a knowledge framework, built over years and fitted to each client. Bret Starr’s system for B2B content rests on decades in the field and years of building. It is doable, but it is not early, and nothing in this repo is close.
Keep an eye out for my next piece, "282 AI Agents Walk Into an Agency's Bar," where we watch them try to run one. (Hint: none of them asks who the client is.)
Want to learn more about this and how I can help you with your AI strategy? Set up a quick chat or shoot me a note.
SHARE THIS WITH YOUR PEOPLE...or with another agency leader who keeps telling you that AI can run the whole agency now.
Figuring out which parts of your agency can be written down, and which can't, is structure work, which is the work I do. I've spent 15 years inside 200+ agencies mapping how the work actually gets done. If you're not sure where that line runs in your own shop, that's worth a conversation.
