Specialist vs Generalist AI Agents - When to Split Responsibilities
Specialist vs Generalist AI Agents - When to Split Responsibilities
One fire did six things. One generalist agent handled email triage, calendar management, file organization, social media posting, code review, and meeting notes. It did all of them adequately and none of them well.
Replacing it with six specialist agents - each focused on a single responsibility - improved reliability across the board. The email agent understood inbox patterns deeply. The calendar agent handled timezone edge cases. The code review agent had the full context window available for diffs instead of sharing it with five other tasks.
The Generalist Trap
Generalist agents are appealing because they are simple to manage. One configuration, one memory file, one set of permissions. But they hit a ceiling quickly.
The context window is the fundamental constraint. A generalist agent needs to hold context for six different domains simultaneously. That means each domain gets roughly one-sixth of the available context. The specialist agent gets the entire context window for its single domain.
This matters more than most people realize. An email triage agent with 200k tokens of context can understand your communication patterns, remember sender preferences, and maintain threading history. The same agent sharing context with five other tasks has 33k tokens for email - barely enough for the current inbox.
When to Split
Split when the domains have minimal overlap. Email triage and code review share almost no context. Running them as separate agents costs nothing in duplicated understanding and gains everything in focused attention.
Keep them combined when the domains are deeply interconnected. Calendar management and meeting notes share context about your schedule, attendees, and priorities. Splitting them means synchronizing state between two agents, which adds complexity.
The Coordination Cost
Specialist agents need coordination. The email agent finds a meeting request and needs to tell the calendar agent. This handoff has a cost - context must be serialized, passed, and deserialized. If coordination costs exceed the benefits of specialization, keep the agents combined.
The sweet spot is usually three to five specialist agents with clear boundaries and minimal coordination needs.
- Focused vs Distributed AI Agent Architecture
- Single Agent First vs Multi Agent
- Five Agents Same Codebase Coordination Lessons
Fazm is an open source macOS AI agent. Open source on GitHub.