Routing Deep Dive

How Routing Works

When a request comes in, Crucinger follows a structured pipeline: classify → select → package → dispatch → track.

Classification

The system analyzes the request to determine domain (engineering? research? content?), complexity (quick answer vs. deep project), and urgency (immediate vs. can wait). This classification drives all downstream decisions.

Agent Selection

Based on the classification, Crucinger selects the best agent for the job. Selection considers agent specialization, current workload, and historical performance on similar tasks.

Context Packaging

Before dispatch, Crucinger bundles relevant context: the original request, relevant memory entries, constraints from the user or project, and success criteria. This ensures the receiving agent can start effectively.

Multi-Step Workflows

Complex requests get broken into sub-tasks. Crucinger sequences the work, manages dependencies between steps, routes each piece to the right agent, and synthesizes results into a coherent output.

Thompson Sampling

Routing improves over time using a Thompson Sampling approach. The system maintains beliefs about which agents perform best on which types of tasks, and updates those beliefs with each outcome. This balances exploitation (routing to proven performers) with exploration (trying new assignments that might be better).