Multi-Agent Architecture

This lesson is adapted from Claire Vo's guide on Lenny's Newsletter.

Claire Vo runs nine specialized agents, each handling a distinct domain of her professional and personal life. This lesson covers her multi-agent architecture, why specialization outperforms generalization, and how to design your own agent team.

Claire's Nine Agents

Agent	Role	Key Tools	Core Tasks
Polly	Personal Assistant	Email, Calendar, Linear	Morning digests, meeting prep, task coordination
Finn	Family Manager	Family calendars, contacts	Family logistics, appointment scheduling, weekend planning
Max	Marketer	Web search, social APIs	Trend analysis, content drafting, website updates
Sam	Sales	CRM, email, web search	Lead qualification, pipeline review, prospect outreach
Holly	Support	Helpdesk, email, Linear	Ticket management, helpdesk automation, inbox organization
Sage	Course Operator	Repos, email, web search	Repository management, course launches, student research
Howie	Podcast Producer	YouTube, email, social	Guest pipeline coordination, episode management
Kelly	Developer	GitHub, code tools	Code generation, PR creation from assigned tasks
Q	Educational	Web search, content tools	Children's educational content, daily lessons

Why Specialization Works

A single all-purpose agent sounds simpler, but specialized agents outperform generalists for several reasons:

Focused context windows. Each agent's AGENTS.md and SOUL.md contain only the knowledge relevant to its domain. A sales agent doesn't need to know about your kids' school schedule, and a family manager doesn't need CRM access.

Clearer boundaries. When an agent has a narrow scope, it's easier to define what it should and shouldn't do. "Holly handles support tickets" is a clearer boundary than "the agent handles everything."

Better tool access control. Each agent only has the API keys and permissions it needs. Your developer agent has GitHub access but not your email. Your family manager has calendar access but not your CRM.

Reduced error surface. When something goes wrong, you know which agent to debug. A sales email sent at the wrong time? Check Sam. A calendar conflict? Check Finn.

Designing Your Agent Team

Map your domains

List the distinct areas of your life/work that could benefit from automation:

Personal (email, calendar, tasks)
Family (schedules, logistics)
Work function 1 (sales, marketing, support, etc.)
Work function 2
Creative projects
Development/technical

Define agent responsibilities

For each domain, write a one-sentence job description:

"Polly manages my email, calendar, and daily task prioritization"
"Finn coordinates family schedules and logistics"
"Sam qualifies leads and manages the sales pipeline"

If a domain is too broad for one sentence, split it into two agents.

Assign tools and permissions

Each agent should have the minimum set of tools it needs:

Agent	Needs	Doesn't Need
Personal Assistant	Email, Calendar, Linear	GitHub, CRM
Developer	GitHub, Code tools	Email, Calendar
Sales	CRM, Email, Web search	GitHub, Family calendar

Set up communication patterns

Agents need to be able to share relevant information:

Shared workspace files in Obsidian or a docs folder
Scheduled handoffs — e.g., Polly's morning digest includes inputs from other agents
Explicit boundaries — agents should know which other agents exist and when to defer

Starting Small

You don't need nine agents on day one. Claire's recommendation:

Week 1: Start with one personal assistant agent (like Polly)
Week 2-3: Add a second agent for your most time-consuming work domain
Month 2: Add agents as you identify clear, repeatable needs
Ongoing: Retire or merge agents that aren't pulling their weight

Each new agent should solve a specific, recurring problem. If you can't articulate the problem in one sentence, you're not ready for that agent yet.

Inter-Agent Coordination

Claire's agents coordinate through:

Shared Obsidian workspace — a markdown-based documentation space where agents read and write information
Morning digest pattern — Polly (personal assistant) aggregates inputs from other agents into a daily briefing
Linear integration — agents create and assign tasks to each other via the project management system
Explicit handoff rules — defined in SOUL.md, e.g., "If a support ticket involves a billing question, flag it for Sam (Sales)"

The key insight: agents don't need to talk to each other directly. They coordinate through shared state (documents, task boards, calendars) — the same way a well-organized human team works.