OpenSquilla: Token-Efficient Agent = Models + Routing Harness

Abstract

OpenSquilla is an open-source, agent-native learnable harness built on one premise: Token-Efficient Agent = Models + Routing Harness — agent capability comes from allocating a pool of heterogeneous models, not from binding to a single strong one. On end-to-end agent benchmarks, its local step-level router preserves 99.96% of fixed-flagship task quality while cutting cost by 88.9%, and its multi-model ensemble routing outscores the strongest single-model baseline (Fable 5) on deep-research tasks at 31% of that baseline's cost. Underneath is a microkernel design: model pools, tools, memory, skills, channels, schedulers, and sandbox policies all plug in through standardized boundaries, separating a stable execution mechanism from evolvable policy. OpenSquilla offers three core competitive advantages. (1) Singleton routing cuts cost: model selection is formalized as a step-level routing operator conditioned on the harness state; in the cost-effective mode the router selects the cheapest sufficiently capable model for each execution step, coupled with runtime mechanisms such as reasoning-budget derivation, prompt caching, and on-demand skill injection, reducing end-to-end cost severalfold while preserving task quality almost unchanged. (2) Multi-model ensemble routing improves quality: in the high-accuracy mode the router assembles complementary model sets and fuses their candidate outputs, exceeding strong single-model baselines in task quality at lower cost and pushing the quality–cost frontier on deep-research tasks. (3) System mechanisms for a learnable harness: the Meta-Skill orchestration layer teaches the agent to retrieve, filter, compose, and evolve skills, distilling user collaboration history into reviewable new workflows; the persistent memory system maintains an editable, retrievable, forgettable, and auditable long-term context for cross-session collaboration; and security governance draws the boundary for real tool execution with tiered policies, runtime isolation, and an audit loop. Together they provide capability orchestration, cross-session state, and execution boundaries, letting the system improve with use while remaining reviewable and controllable. Comparative analysis shows that the value of OpenSquilla lies not in stacking a single large model, but in raising the agent capability obtainable per unit cost through agentic routing (singleton routing for cost, multi-model ensemble routing for quality), Meta-Skill orchestration, long-term memory, and governed tool boundaries.

Keywords