Tool Specialization

When systems accumulate many tools or actions, generalist reasoning becomes
unreliable. The Tool Specialization pattern isolates tool usage behind narrowly
scoped reasoning units, reducing decision complexity while preserving capability.

This pattern captures invariants and tradeoffs without prescribing a specific
implementation.

This pattern emerges when adding tools increases confusion, misfires, or decision
paralysis instead of effectiveness.

Typical triggers:

• a growing tool menu that the "main" agent uses inconsistently
• repeated tool misuse despite prompt reminders
• correct tools, wrong selection timing
• tool choice dominates reasoning time

Assign responsibility for a narrow set of tools or actions to a dedicated reasoning
unit with constrained scope and explicit outputs.

The goal is not "smarter" reasoning.
The goal is making tool-use boring and reliable.

• Capability grows faster than reliability without isolation
• Isolation precedes orchestration
• Specialization reduces decision load, not intelligence
• Outputs must be explicit and promotable

• Increased coordination overhead
• Additional interfaces to maintain
• Risk of premature specialization if created before pressure is visible

This pattern does not define:

• an agent framework
• a required topology
• how sub-agents are implemented in any specific repo

• Cognitive Partitioning — Universal concept
• Canonical Compression — Reduce reasoning surface area (context limits)