Eric Siu published a 12-point breakdown this week of how his agent runs his business. He calls it his chief of staff. It books sales calls, revives stalled deals, and ships ad creative while he is at the gym. The piece is full of useful specifics: Linear for the spine, Obsidian for memory, OAuth instead of metered tokens, Tailscale to keep the agent reachable from any device. Worth your time.
But the headline insight is not the tool list. It is one line near the end.
"You have to give up doing the task and start designing the system that does it."
Most people are not going to do that. Eric says so himself in the same paragraph. They will use the agent as a fancy assistant and stop there. The assistant version is comfortable. The operator version asks them to redesign how they work.
That gap is the most important thing happening in agencies right now.
Two versions of the same tool
The assistant and the operator look identical from the outside. Same model. Same chat window. Same API key.
What is different is the relationship.
flowchart LR
A[Assistant mode] --> B[You do the work]
B --> C[Agent helps you do it faster]
C --> D[Output ships when you ship it]
E[Operator mode] --> F[Agent does the work]
F --> G[You design the system that does it]
G --> H[Output ships on a loop, with or without you]
classDef mode fill:#001D22,stroke:#001D22,color:#F5F7F7
classDef step fill:#FFFFFF,stroke:#001D22,color:#001D22
class A,E mode
class B,C,D,F,G,H stepAssistant mode is a speed multiplier on the work you were already doing. Operator mode is a model where the work stops being yours.
Both are real. They are not on the same continuum. Picking one closes the door on the other for as long as you keep picking it.
Assistant mode is where agencies are dying
Most agencies are deep in assistant mode and do not know it.
The tell: the agency uses AI inside every deliverable, but the deliverables are unchanged. The campaign brief is still a Google Doc. The content calendar is still a spreadsheet. The reporting deck is still a slide template. ChatGPT writes the first draft. A human edits. The thing ships. The cycle repeats.
That is the same business that existed in 2018. The only thing that changed is the typing.
The pricing model still bills for hours. The team still grows linearly with client count. The client still gets a monthly report that nobody reads. The agency still loses three weeks every time the account manager turns over. The agency still cannot tell you, on Tuesday at 2pm, what state every active retainer is actually in.
AI made the typing faster. The business is identical. Which means the margin is identical, and the moat is identical, which is none.
By Q4 2027, this version of the agency looks like Mad Libs for ChatGPT. The client figures out they can run the same prompts themselves. The retainer churns. The agency cuts headcount. The slow unwind begins.
Operator mode is a different product
Operator mode is not "more AI." It is a different operating model.
In operator mode, the agency does not sell deliverables. It sells a system that produces outcomes. The deliverables fall out of the system. They are an artifact, not the product.
flowchart TB
subgraph Assistant[Assistant-mode agency]
A1[Client brief] --> A2[Team drafts with AI help]
A2 --> A3[QA and send]
A3 --> A4[Monthly report]
A4 --> A1
end
subgraph Operator[Operator-mode agency]
O1[Client signal] --> O2[Agent triages]
O2 --> O3[Specialist agents execute]
O3 --> O4[Human reviews and approves]
O4 --> O5[System logs outcome]
O5 --> O6[Loop learns from outcome]
O6 --> O1
end
classDef step fill:#FFFFFF,stroke:#001D22,color:#001D22
classDef accent fill:#F9432B,stroke:#F9432B,color:#FFFFFF
class A1,A2,A3,A4,O1,O2,O3,O5,O6 step
class O4 accentOperator mode requires three things assistant mode does not:
- A loop. Not a workflow. A loop. Trigger to action to verification to memory to next trigger.
- Memory. The agent has to remember what it did yesterday and what worked. Eric uses Obsidian and a setup called QMD. The point is the same: built-in model memory is not enough.
- A human approval gate that ships at the speed of the work, not at the speed of meetings.
Once those three are in, the work shifts. The operator stops doing tasks and starts watching outcomes. The team stops growing linearly with client count and starts growing with system count.
The five shifts
What changes when you cross from assistant to operator. None of these are tool decisions. All of them are operating decisions.
- From documents to loops. The unit of work stops being a deliverable. It becomes a loop with a trigger, an action, a verification step, and a memory write.
- From specialists to manager agents. You do not hire one agent to do everything. You hire a manager agent that coordinates specialist agents. Eric calls this "an army of one." Same idea.
- From hours to outcomes. Pricing follows. If the loop ships the outcome, hours are noise. The retainer prices the outcome.
- From quarterly reports to live state. The system always knows the state of every account. Reporting becomes a query, not a project.
- From meetings to approvals. Most internal meetings are status. If the system surfaces status, the meetings disappear. What is left is decisions, and those happen in approval gates measured in minutes.
Cross all five and the business looks nothing like an agency anymore. It looks like a product company that happens to deliver services as the output.
The trade is brutal at the start
Eric is honest about this. The first stretch is messy. You will babysit the agent more than it saves you. You will rip out half of what you set up.
That is correct, and it is the reason most operators give up at month two.
The first 90 days of operator mode look like a step backward. The loops break. The memory file corrupts. The agent posts something it should not have. The team complains that the new system is slower than the old way. You spend evenings cleaning up.
Then somewhere around day 100, the loops start compounding. The agent stops needing as much correction. The memory starts shaping the next draft before you do. A specialist agent ships a piece of work end to end without you watching, and the output is better than the version you would have shipped manually at 11pm on a Tuesday.
That is the moment. After that, the curve bends.
The agencies that quit at week six never see the curve. They report back that "AI did not really work for us" and quietly return to assistant mode. The ones that push through emerge with a business that scales without headcount.
What this looks like at Magnet
Concrete, because vague does not help anyone.
Magnet's content engine is in operator mode. The loop drafts the day's longform, scores prior posts after three days, biases the next draft toward what worked, cross-posts to the blog as MDX, and writes the campaign row to Notion. It runs five days a week. The human approval gate is one Slack DM per piece. Pre-approval is not approval. Same-day approval is the only gate that ships.
Magnet's PM function is in operator mode. Twice a day on weekdays, an orchestrator reads every client signal across Gmail, Slack, Linear, GitHub, and Calendar, files it into the right initiative, drafts client replies, and writes a Linear project update on every active project. Status surfacing happens through Linear Pulse, not Slack. The PM cron does the work a PM used to spend half their day doing, and the human approves the parts that need human judgment.
Magnet's finance function is in operator mode. One skill owns the books. Weekly card sweeps, monthly Chase reconciliation, AR sync, overdue invoice chase, pass-through billing. Five crons across the week. Slack DM on anything that needs human eyes.
None of these are deliverables. They are systems that produce deliverables.
Magnet's headcount has not grown to support any of them. That is the point.
The choice
Eric's piece is a good map. The 12 jobs he describes are real. The stack he uses works. You can copy most of it.
The harder thing he is asking you to do is the operating shift. Stop being the person who does the task. Start being the person who designs the system that does it. That trade is the actual product.
If you are a founder, the question to sit with is whether you are willing to feel slower for 90 days to be 10x faster forever. Most founders are not. The few who are will own the next decade of their categories.
If you are an agency principal, the question is sharper. You either rebuild the agency on operator mode or you become the next quarter's case study in why some agencies could not make the transition. There is not a middle path that holds for very long.
The assistant is comfortable. The operator is the prize.
Work with Magnet
If you are an agency principal or a founder running AI in assistant mode and you can feel the ceiling, Magnet runs a Loop Audit for operators making this transition. We sit down with your current ops, identify the three highest-leverage loops to build first, and ship the first one inside 30 days. Fixed price. No retainer required.
Five slots this quarter. Book one at https://magnet.co/contact. Closes when filled.