How We Structure an AI Agent Fleet for a Real Agency
Running a coordinated fleet of specialized AI agents sounds futuristic until you have to actually build and manage one. Here is what the architecture looks like at Merlino AI today — the roles, the handoffs, and where it breaks.
Why a fleet, not a single agent
The temptation is to build one super-agent that handles everything. We tried that. The context window fills up, the instructions contradict each other, and the outputs become unpredictable past a certain task complexity. The better model is narrowly specialized agents that hand off to each other cleanly.
The current roster
Our fleet currently covers five lanes:
- Sales: Outreach sequencing, follow-up logic, objection handling frameworks
- Content: Research, drafting, rewriting, headline testing
- Dev: Front-end builds, component generation, deployment scripts
- SEO: Keyword research, content briefs, citation tracking, audit reporting
- Ops: Project management, client communication drafts, internal SOPs
Each lane has a primary agent and a set of tools it can call. Agents do not cross lanes without a handoff protocol.
The handoff problem
The hardest part of multi-agent orchestration is not the AI — it is the handoff. When does Sales hand to Ops? What context does it pass? What happens when the receiving agent does not have enough information to complete the task?
We solved this with structured context packets: a standardized JSON format that every agent knows how to read and write. When an agent completes its phase of a task, it writes a context packet and triggers the next agent. The receiving agent reads the packet before it touches anything else.
Where it still breaks
Ambiguous briefs. If the initial input from a client or team member is vague, the first agent in the chain makes assumptions, and those assumptions propagate through every downstream agent. Garbage in, compounded garbage out. The fix is a human-in-the-loop intake step before the first agent fires — something we added after a painful client delivery.
What this looks like day to day
On a typical client campaign, the fleet runs mostly unattended. I review outputs at key checkpoints: after research is done, after the first draft is out, and before anything goes to the client. The agents handle the volume; I handle the judgment calls.
That is what AI-native actually means. Not replacing human judgment — delegating the repetitive execution so human judgment is spent where it actually matters.