FG-0301 Agent Specialization
Description
Design AI agents with distinct roles, domains, and capabilities rather than using generalist agents for all tasks. A single generalist agent produces generalist results: competent but unfocused. Specialized agents with defined domains produce deeper, more consistent output because their entire context window is dedicated to their specialty. This is the multi-agent design pattern, and at the system level it echoes mixture-of-experts (and, more directly, multi-agent routing): route each task to the specialist best suited to it. Specialization also reduces contradictory requirements: a testing agent does not need to balance quality advocacy with shipping pressure.
Implementation
Define 3-10 agent roles based on your team's actual workflow domains (not theoretical). Each agent should have: a clear domain, defined capabilities, documented limitations, and a distinct operational profile. Start with the minimum viable number of agents and add specialists when a generalist consistently produces shallow results in a specific domain.
Success Indicators
- Each agent produces domain-expert-level output in their specialty
- Task routing is clear: the team knows which agent handles what
- Agents don't produce contradictory outputs due to conflicting requirements
- Specialized output quality measurably exceeds generalist output in the same domain
Failure Modes
- Too many specialists: coordination overhead exceeds specialization benefit
- Specialists too narrow: gaps between domains go unaddressed
- Specialization without documentation: team doesn't know agent capabilities
- Artificial specialization: creating roles for the sake of structure rather than need
Sub-Methods 3
Create a formal profile document for each specialized agent that defines: primary domain, secondary capabilities, explicit limitations, interaction preferences, and escalation triggers. The document serves as both onboarding material and operational reference. It should be loaded during context recovery.
Establish clear boundaries between agent domains to prevent overlap conflicts and ensure coverage gaps are identified. When two agents could reasonably handle a task, define primary and secondary ownership. When no agent covers a domain, flag it as a coverage gap rather than letting a generalist fill it poorly.
Monitor and measure how deeply each agent develops expertise in their domain over time. Track metrics like output quality ratings, domain-specific vocabulary usage, and consistency of domain-appropriate decisions. Agents whose specialization plateaus may need profile refinement or domain scope adjustment.