Capability-class routing

ALDO AI is LLM-agnostic by charter. No agent names a vendor model. Instead, every agent declares a capability class — a coarse-grained tier of reasoning power and budget — and the gateway resolves a concrete model at call time, taking into account the tenant's catalog, the agent's privacy tier, and the agent's budget.

#The classes

There are five classes shipped out of the box:

• reasoning-large — frontier models for hard, multi-step reasoning.
• reasoning-medium — strong general-purpose reasoning at lower cost.
• reasoning-small — fast, cheap, good for routing or extraction.
• local-reasoning — runs on the operator's hardware. Always allowed for sensitive privacy tier.
• fast-draft — for streaming first-token latency.

A spec picks one:

model_policy:
  capability_class: reasoning-medium
  fallback: local-reasoning

#Why classes, not model names

If an agent spec hard-codes gpt-4o, three things break:

1. The spec leaks a vendor decision into the agent layer. A different tenant with a different licensing posture can't run that spec.
2. Local mode is impossible. There's no way to swap the cloud model for an Apple Silicon checkpoint without rewriting the spec.
3. Eval comparisons (cloud vs local on the same agent) require code changes, not config changes.

Capability classes invert that: the agent declares the requirement, not the fulfilment. Swapping a tenant from cloud to local-only is a single change to the routing config, applied to every agent at once.

#Routing inputs

The router considers, in order:

1. Privacy tier. sensitive runs MUST hit a local provider. Anything else is dropped before a request leaves the box.
2. Capability class. Filter the catalog to models that satisfy the requested class.
3. Budget. Drop models whose unit cost would exceed the run's per-step budget.
4. Locality preference. When two models pass, prefer the one marked locality: local.
5. Health. Skip models the gateway has marked unhealthy in the last minute.

If no model survives, the request fails fast with a structured error the run UI surfaces clearly.

#Inspecting decisions

Every replayed run includes the routing decision per step in the event log. The run detail page renders it under Routing → Decision with the candidate set, the selected model, and the reason any candidates were dropped.

#Related

• Privacy tier
• Replayability