Agents & Cognition · 06.03
Model Routing: Minimum-Sufficient Tiers
How each task is classified and routed to the cheapest sufficient model tier, unless an explicit pin or override wins.
Six task classes
Every routing decision starts by classifying the task: trivial, simple_text, standard_reasoning, workflow_synthesis, tool_heavy, or high_risk. Classification is signal-based and ordered — a high-risk signal wins outright regardless of length; required affordances or tool-heavy language route to tool_heavy; short, simple language only counts as trivial under a length ceiling, so a short but genuinely risky request never slips through as trivial just because it's brief.
Nine possible sources for one decision
A routing decision reports exactly where it came from: explicit_pin, workspace_role, workspace_default, env_role, env_default, runtime_detected, agent_config, fallback, or custom. An explicit pin always wins outright — the minimum-sufficient-tier logic only ever runs when nothing more specific already decided the model. Every decision also carries its rejected alternatives with reasons, so "why did it pick this model" is always answerable, not just the final choice in isolation.
Continue
Every workspace agent is ranked ready, offline-capable, enablable, or incapable against a node's declared requirements — never a hardcoded agent name.
Sessions persist across turns, but the prompt is reconstructed each time — persona, task, plan, and fresh observations reassembled, not resent stale.