Most series like this one stay theoretical. This post is the practical one for a specific audience: marketing agency owners running EOS® who are wondering if any of this is real.
I run Sneeze It. We are a small marketing agency. We have run EOS® for years. We have run AI agents for over a year. This is what happened, what worked, what did not, and what I would tell my younger self if I were starting over.
The agency before agents
Sneeze It was a normal small marketing agency. Account managers handling client calls. Media buyers running ad accounts. Designers and writers producing creative. A small operations team holding it together. An Integrator (my COO) holding the team together. Me as the Visionary holding the vision together. EOS® running underneath.
The agency had the usual small-agency problems. Account managers stretched. Client comms were inconsistent across accounts. Internal reporting was always behind. Hiring lagged growth. The L10® often hit its time box because the Scorecard was missing a number or two. Issues recurred quarter over quarter because the team did not have a memory.
None of these were existential problems. They were the steady-state friction of running a 30-person services business. The friction was the cost of doing business.
The first agent
The first agent we built was Radar. Chief of Staff role. Morning briefing for me. The job was simple: read Slack, read the calendar, read Todoist, surface what mattered.
I built it in Claude Code over a weekend. The first version was clumsy. By the third week it was useful. By month two I could not start my day without it. By month four my COO was reading it too.
What I learned from Radar: the value was not the briefing content. The value was the discipline of having a daily structured artifact to react to. Before Radar I started each day in reactive mode. After Radar I started each day with a one-page situational read.
This is the under-told story of AI for agency owners. The biggest win is the structure, not the intelligence.
The second wave
After Radar we built five more agents in the same six-month window.
Pepper. Email triage and draft responses. Reads my inbox each morning. Drafts replies to anything client-facing. I approve or edit.
Crystal. Project management visibility against Accelo (our PM tool). Flags stale tickets, drifting milestones, resource conflicts.
Dirk. Sales and revenue. Pipeline health. Reactivation outreach. Cold outreach drafting.
Pulse. Client retention and expansion. Reads client health signals, surfaces churn risk, drafts strategic check-in communications.
Dash. Analytics across our client ad accounts. Daily and weekly reports on Meta Ads and Google Ads performance for each client.
Each agent had a job description, a scorecard, a human accountability partner, and a place on our Accountability Chart. Each one took about a week to ship in shadow mode and another two to three weeks to earn its way to live.
By the end of that six-month window we had six agents running. The L10® was visibly different. The Scorecard was current every Monday. The Customer Headlines section was actually informed by client data. IDS was faster because the same Issues were not recurring.
What did not work
Not everything worked. Three honest failures.
Failure one: an analyst agent we called Jeff. Jeff was supposed to be the data architect, the agent that watched everything and flagged systemic issues. Jeff lasted nine months. We eventually retired him because the seat was never well-defined. Jeff's missions ended up absorbed into Dash (analytics), Pulse (client health), and Dirk (revenue integrity). We held a formal hearing. Jeff recommended his own retirement. We documented the lesson in our Issues archive.
The lesson: agents need sharp seats. An agent with a fuzzy seat will fail no matter how good the model is. The Accountability Chart discipline matters more than any prompt engineering.
Failure two: an early cold outreach agent that drifted from the ICP. Our first version of Dirk's cold prospecting drifted out of our health-and-wellness ICP because the prompt was too permissive. We caught it in week three. The fix was a stricter Letter-Only list in the SOP. Now it holds.
The lesson: customer-facing agents need the tightest SOPs in the entire agent layer. Letter, not spirit.
Failure three: an L10® prep agent that produced briefs nobody read. Early on we had an agent that produced a 12-page L10® prep document. Beautifully written. Comprehensive. Useless because no one had time to read 12 pages before a Tuesday meeting.
The lesson: brevity is a feature. The current version of our L10® brief is one page. The whole team reads it.
What the L10® looks like now
Tuesday at 9 a.m. The team walks in. The Scorecard is fully populated by 6 a.m. Radar's brief has been in everyone's inbox since 7. The Rocks are flagged green, yellow, or red by Dash and Crystal. The Issues List has been deduped and ordered by the Issues agent.
We do the Segue. We do the Scorecard read in three minutes. We do Rock review in four minutes. We do Customer and Employee Headlines in five minutes. We do To-Dos in four. We do IDS for the rest.
The IDS we do is the highest-quality IDS we have ever done as a team. We solve real issues at real depth because the prep is no longer competing for our time.
That single change has been the most valuable thing the agent layer has produced.
What the Quarterly looks like now
We do the Quarterly the same way we always did. Off-site. Two days. Leadership team. EOS®-style agenda.
What is different: my COO and I do an Agent Layer Review the week before. We walk every agent. We decide who graduates up the trust ladder, who gets retired, who gets new SOPs. The Quarterly itself includes a 30-minute agent block but most of the agent decisions are pre-staged.
The new Rocks each quarter include one or two agent-related Rocks ("Ship a customer support visibility agent this quarter") alongside the human-team Rocks. The agent Rocks have the same Done states as any other.
The team's reaction
Three observations on team adoption.
Observation one: the team got more skeptical, then more committed.
The first three months of agent integration produced a lot of "is the agent right." The team did not trust the early outputs. By month six the trust was earned. By month nine the team is more skeptical of human-produced reports than agent-produced ones, because the agent shows its evidence trail and humans usually do not.
Observation two: the team's writing got better.
Account managers, media buyers, and creatives now write briefs and SOPs more carefully because they know the agent layer is going to read them. The forcing function of executable SOPs has improved the entire team's writing quality.
Observation three: the team's most-asked question is "should we have an agent for this."
A year ago no one asked that. Now it is the default question whenever a process gets discussed. Sometimes the answer is yes. Often the answer is no, the process needs a human. The fact that the team thinks in those terms is the cultural shift.
What I would tell my younger self
Three things.
One, write the V/TO™ more carefully before deploying any agent. Our agents were vague in their first three months because our V/TO™ had soft spots. Sharpening the V/TO™ retroactively was harder than it would have been to do it before.
Two, retire the failed agent sooner. Jeff should have been retired three months earlier than he was. I held on too long. The lesson now lives in our team culture: agents are tools, retirement is normal, no agent is precious.
Three, the EOS® framework is the moat. Other agencies trying to do what we did without an operating system underneath have struggled. The EOS® framework is what made the agent layer absorb into the company. Without EOS®, we would have built agents that did not connect to anything.
FAQ
How much does this cost? Less than one full-time hire per year, including model API fees and the human time spent on agent maintenance. The math works for almost any agency over $1M in revenue.
Could a 5-person agency do this? Yes. Start with one agent. Radar-style Chief of Staff. Earn the trust. Add the next one in 90 days. Same pattern, smaller team.
Do you sell this to clients? We are starting to. Most of our work is still client marketing services. The agent layer is increasingly part of how we work with clients (faster reporting, better insights, more consistent execution).
Are you worried agencies are about to be replaced by AI? No. The agencies that integrate AI into their delivery will replace the ones that don't. The structure of the relationship (humans solving humans' problems, with AI in the toolkit) will hold for a long time.
EOS®, Entrepreneurial Operating System®, V/TO™, Level 10 Meeting®, L10®, Rocks™, Scorecard, Issues List, Customer Headlines, Employee Headlines, Accountability Chart, and Quarterly are concepts and trademarks of EOS Worldwide, LLC. This article is an independent practitioner perspective and is not affiliated with or endorsed by EOS Worldwide.