Specialist or Generalist Artist
Specialist or Generalist Artist
A coding agent that also handles email, schedules meetings, and writes blog posts will be mediocre at all of them. A coding agent that only writes code will be excellent at it. The question is whether excellence in one area is worth blindness in everything else.
The Case for Specialists
Specialized agents carry less context, make fewer errors in their domain, and can be optimized aggressively for one task type. A code review agent does not need to know how to send emails. Stripping that capability away means:
- Smaller system prompts - less context to confuse the model
- Targeted tools - only the tools relevant to the task
- Better evaluation - you can measure code review quality precisely
- Simpler failure modes - fewer things to go wrong
The Case for Generalists
Generalist agents handle the unexpected. Real work does not come in neat categories. A task that starts as "review this PR" might require reading a Slack thread, checking a dashboard, and sending a summary email. A specialist agent stops at the boundary of its expertise. A generalist pushes through.
The Right Answer Is Both
The pattern that works is specialist agents coordinated by a generalist orchestrator:
- The orchestrator understands the full workflow and breaks it into tasks
- Specialist agents handle each task with domain expertise
- The orchestrator reassembles the results and handles cross-domain logic
This gives you depth where it matters and breadth where it is needed. One agent does not try to be everything.
The artist analogy holds. A studio has specialists - colorists, animators, sound designers. But it needs a director who understands all the disciplines enough to coordinate them.
Fazm is an open source macOS AI agent. Open source on GitHub.