SID-2 Swarm

Overview

How a SID-2 Swarm run works: a team of agents searches your data in parallel, and what changes from SID-2.

SID-2 Swarm is the multi-agent version of SID-2. It answers the same kind of question and returns the same kind of result: ranked IDs for the documents needed to answer it. Instead of one agent searching turn by turn, a root agent forks itself into many agents that search in parallel, message each other about the documents they find, and each report their own ranking. We recommend to fuse rankings returned by agents using Reciprocal Rank Fusion (RRF).

Single agent vs swarm

A single SID-2 episode is one agent: it searches, reads, adapts and reports. A SID-2-swarm runs as a team of agents working on one question at the same time:

  • The run starts with one agent, root
  • Each agent can call the fork tool with a list of names, one per new agent
  • Each forked agent starts from its parent's transcript so far. From there every agent works independently, with its own turns.
  • Agents send each other messages by name. An agent can read any document another agent found.
  • Each agent ends by calling report_helpful_ids. You fuse the reports into one ranking, for example with reciprocal rank fusion (RRF).

Swarms help most on questions with parts that can be researched separately, such as multi-hop and comparison questions. For simple questions a single SID-2 episode is cheaper.

What stays the same

You still own your data and the code around the model. As with SID-2:

  • You implement the search and read tools against your backend, with the same tool schemas and the same fixed report_helpful_ids tool.
  • You render results with the SDK's DocumentCache, as described in Formatting documents.
  • The instructions, tool_examples and reward mean the same thing.

What is different

  • SID handles orchestrating fork, message, wait and list_agents calls, and adds the multi-agent protocol to your instructions.
  • Reasoning effort is fixed at high; you don't set it.
  • You handle searches and reports asynchronously, as they arrive. You send one request and keep a WebSocket open. Each search or read call is tagged with the agent that made it. You execute the tool call and send the result back on the socket.
  • When an agent calls report_helpful_ids, you tell SID to stop that agent.
  • When an agent forks, you fork its DocumentCache with DocumentCache.fork(n). The DocumentCache ensures that document ids are consistent between agents, so that agents can read documents based on ids sent by other agents, and that documents seen before fork is called are marked as seen.
  • Currently, SID-2-Swarm is only available via SID's first-party api, api.sid-1.com.

Content