How to Build Team Agents
Summary
In this Operator's Cut episode, host and guest Nofar Gaspary discuss the shift from solo AI agents to "team agents" that operate within shared workspaces. They define four archetypes—expert, common work, bridge, and chief of staff—and outline five core design decisions regarding scope, hosting, knowledge, permissions, and ownership. The conversation highlights why not every agent should be shared, identifying scenarios where private agents remain superior, and provides a practical playbook for building effective collaborative AI tools.
Topics discussed
Introduction: The shift from individual to team AI agents
Sponsors and upcoming webinar on personal AI benchmarks
Defining team agents and the need for intentional design
Motivating stories: Knowledge silos and the 'agent ball'
Terminology: Multiplayer AI, shared agents, and AI mates
KPMG study: Skills vs. outcomes in AI-assisted work
Sponsor segments: Blitzy, Harbor Capital, and Hyperagent
The spectrum: Private agents vs. shared knowledge vs. team agents
Archetype 1: The Expert Agent (centralized know-how)
Archetypes 2 & 3: Common Work and Bridge Agents
When NOT to build a team agent: Ownership and complexity
Case study: Defining scope and 'do not' lists for a customer agent
Deployment options: Shared folders, vendor tools, and custom harnesses
Privacy and memory: Who sees what and where learning is stored
Knowledge management: Collecting, verifying, and maintaining data
Security rules: Access control and preventing data leakage
Operations: Ownership, rules of engagement, and monitoring
Conclusion: Course promotions and future predictions
Q&A: Common starting points and building vs. waiting for tools
Listen ad-free on Castria