Capability Tiering

Capability tiering is the practice of assigning different models and tool sets to different agent roles within a multi-agent system, forced by the economic cost of inference rather than by architectural preference.

Forced-by constraint

Inference cost scales with model size. Running a frontier model on every subtask in an agent pipeline is economically unviable at production throughput. Capability tiering is the structural response: match model capability to task complexity, and restrict expensive models to roles where their reasoning ceiling is actually needed. The constraint is not optional -- it is imposed by the unit economics of token-based billing.

Members of the class

Changelog-derived evidence places three tiers in active use in this system:

• Haiku-class workers -- high-volume, low-complexity subtasks; pushed hard for token efficiency.
• Sonnet-class workers -- mid-complexity synthesis and execution tasks; the default workhorse tier.
• Opus-class synthesis -- reserved for final synthesis, adjudication, or decisions where reasoning depth is the bottleneck.

Tool access follows the same gradient. Workers in lower tiers receive narrower tool sets matched to their defined scope; senior tiers receive broader authority. This pairing of model tier with tool tier is the defining feature of capability tiering as a primitive class.

Why it matters

Tiering is not merely cost optimisation. It enforces accountability boundaries: a worker operating below its capability ceiling cannot accidentally make decisions that belong to a higher tier. It also makes orchestration legible -- the tier of a delegate signals its expected scope before any output arrives. Agent fields are described as instances of this primitive, meaning a deployed agent field concretises the abstract tiering policy into specific model assignments and tool grants for each role.

Caveats

This article is changelog-derived and has not been runtime-verified. The tier labels (Haiku, Sonnet, Opus) reflect a specific provider's model family and may not generalise across providers. The assignment of tools to tiers is asserted as a pattern, not as a measured outcome. Whether tiering produces the intended accountability boundaries in practice -- or whether cost savings are offset by coordination overhead at tier boundaries -- remains untested in this record.