When AI Agents Work in Teams
Multi-agent workflows create distributed accountability that challenges existing legal frameworks designed for single AI tools rather than collaborative AI systems with emergent behaviors and shared decision-making.
Engineers are working with teams of specialized AI agents rather than single coding assistants. Each agent handles specific tasks—such as code generation, testing, security review, and documentation—while collaborating through shared knowledge bases and handoffs between them. Legal teams need to rethink their approach to governing these systems.
We're moving from governing individual AI decisions to governing AI team dynamics. The challenge isn't just harder—it's a different category of problem. When specialized agents collaborate through parallel workflows and shared context, responsibility gets distributed across the system rather than concentrated in individual tools.
What does this look like? An agent team deploys code that later causes data exposure. Traditional incident response asks which model made the problematic decision. But in multi-agent workflows, the coding agent might have written secure logic, the testing agent might have followed proper protocols, and the security agent might have correctly applied existing policies. The problem could emerge from how they interact rather than from individual failures.
Current AI policies can't handle this. They focus on overseeing individual agents—defining boundaries for specific tools, establishing review processes for discrete outputs, and assigning responsibility for single-point decisions. Multi-agent systems need different approaches designed for distributed decision-making where outcomes emerge from collective action.
Development teams are already implementing these workflows through platforms like Claude Code and Warp, often without waiting for policies to catch up. Legal teams that adapt quickly will shape how these systems develop rather than just responding to problems after they surface.
Bill Doerrfeld
