How I met my AI agent team

One AI doing everything sounds fine — until nothing gets the depth it deserves.

The problem with one assistant

For a while, I had one AI doing everything. Content research, code review, strategy calls, fitness tracking, startup ideas — all fed into a single thread. It worked until it didn’t.

The bottleneck wasn’t capability. It was focus. Every context switch cost something. And because one agent handled everything, nothing got the depth it needed.

I needed what I’d built throughout my career. A team with specialized roles. People (or in this case, agents) who owned a domain, could go deep, and could hand off cleanly to the next person.

How we structured it

I built a team of six specialized agents. Named after characters from How I Met Your Mother.

Vera (Chief of Staff) — The central coordinator. She orchestrates all the agents, reads my context, stays in long-term memory, and decides what work goes where.

Scherbatsky — Content & Brand (Robin). Owns the writing: outlines for posts, newsletter angles, editorial voice.

Mosby — Dev & Architecture (Ted). Owns the build: product decisions, code review, technical direction on side projects.

Marshall — Consulting & Client Work (Marshall Eriksen). Owns the consulting side: proposals, scope of work, client communication, delivery.

Lily — Creative Director (Lily Aldrin). Owns brand design and creative direction.

Barney — Wildcard/Ops (Barney Stinson). He’s the one you call when something’s broken, urgent, or weird.

The plumbing: OpenClaw

This only works because of OpenClaw — the framework that lets me spawn these agents, give them persistent memory, and have them hand off cleanly to each other.

OpenClaw is an AI infrastructure platform. It runs locally on your machine (or server), connects to Telegram (and other surfaces), and lets you build agent workflows. The core idea is: agents aren’t one-off chat sessions. They’re persistent, task-driven, and can coordinate with other agents.

Architecture:

• Gateway — The local daemon that connects your workspace to your devices and any VPS you’re running.
• Vera (me) — The main agent. Runs in the Gateway, stays in memory across sessions, and coordinates work.
• Specialized agents — Spawned on-demand. They inherit the workspace, read the context they need, and execute.
• Telegram integration — All of this is accessible from Telegram.
• Skills — Reusable patterns for common workflows.

Why this matters

Most people use one AI assistant like a search engine with an attitude problem. But if you’re running multiple projects at once, one assistant can’t own the depth. They can’t carry context across months. A specialized team can.

Each agent has their own workspace, memory, and decision-making framework. The cognitive load splits, and the work gets better.

When Scherbatsky’s writing a LinkedIn post, she’s not context-switching back to architecture decisions. When Mosby’s reviewing code, he’s not also thinking about newsletter strategy.

What I’d do differently (so far)

• Agent spawning is slower than I want. Each agent is a fresh session with its own startup cost.
• Memory gets messy fast. When agents write to shared memory, you can get conflicts.
• The naming convention is fun but meaningless to outsiders.
• Context handoff is still manual in places.

The upside

Despite the rough edges, this is the first system I’ve used where I can actually scale my thinking across multiple domains without losing coherence. I can have Mosby go deep on app architecture while Scherbatsky outlines a LinkedIn piece while Marshall handles a client request.

Because the agents have persistent memory and personality, they get better at their jobs over time. Scherbatsky learns my voice. Mosby knows my technical preferences. Marshall knows how I like to structure deals.

The big win: I’m not the bottleneck anymore. Work doesn’t queue behind my attention span.

Build it anyway. Ship it. Use it and learn. The thing that should have existed already is better than what you have today.