Router

Using a routing mechanism to distribute requests to models of different scales based on task complexity can reduce the average call cost at the overall level.

The Super Agent architecture uses a two-tier structure of Router Agent and Worker Agent, with the Router Agent serving as the user entry point responsible for coordination and the Worker Agent responsible for executing tasks.

DeepSeek open-sourced a new architecture called Super Agent, which includes two roles: Router (dispatcher) and Worker (executor).

Walkers do not share intermediate information with each other by default, and cross-task information transfer must be relayed through the Router.

DeepSeek open-sourced an Agent prototype that uses a two-layer architecture (Router and Walker).

The Router is responsible for interacting with users, understanding intent and distributing tasks to Walkers, and does not execute specific tasks itself.

The Router of Super Agent is deliberately designed to be unable to execute tasks itself, in order to avoid a dispatcher bottleneck.

In the code, the Router main painting has 14 application tools, but there are no Write, Edit, Blob, or Grab tools, and no Subagent.

SuperAgent adopts a Router-Worker architecture that separates message triggering, task-acceptance decisions, long-term work items and actual execution: the Router decides whether to accept a task and whom to assign it to, and the Worker completes the task in an independent session and directory and writes back the result.

If the Router layer only performs intent dispatching without input sanitization, injected content may reach the Worker directly, so permission gates usually need to be paired with content filtering and pre-execution confirmation.

In dsh-superagent, the Router is responsible for receiving tasks, determining intent and dispatching, while the Worker is responsible for actually executing time-consuming actions such as code generation and file operations.

Router is Layer's routing layer, essentially a lightweight classification/decision model that reads the input text (and optionally State), determines the problem domain and forwards it to a downstream model, and the Routing field in the returned result records the target model and reason.

The gating network is usually a lightweight linear layer that maps a token's hidden state to a score for each expert, and after Softmax takes the top-K experts by score.

For developers who need to run Agents over the long term, the cost savings brought by model routing are usually more significant than referral link discounts.

Router has two typical implementation paths: rule-based heuristics (routing by instruction length, keywords or regular expression matching, with zero latency, zero additional cost and predictable behavior) and model-assisted routing (using a lightweight model to classify intent and then select the downstream model, with higher accuracy but introducing an extra call).

The routing judgment itself also consumes tokens, and when the task itself is very short the routing overhead may take up a larger proportion, so the routing logic should not be made overly complex.

The Router in a MoE architecture dynamically decides which experts to send the input to during each token processing step.

A dual-model strategy typically requires a routing layer (Router) to automatically allocate traffic based on request complexity, response length, or business priority.

Ramp co-founder Eric Glyman was personally involved in building the Router product.

As long as a cheaper model can meet the set quality requirements, requests are automatically routed to it, achieving a dynamic trade-off between cost and quality.