swarm

Create a table from a glob pattern or file list, then call run() with an instruction template containing {file} placeholders and a responseSchema. Each file becomes one row, and rows are dispatched to subagents with bounded concurrency and merged back automatically.

Read and parse the file inside the eval sandbox, pass the records to create({ tasks: records }), then run() with an instruction like "Classify {text}" and a responseSchema defining the label field. One record equals one row; do not group multiple items into a single row.

Omit subagentType for classification, extraction, and labeling where a single direct model call with structured output suffices. Set subagentType when the task needs tools, file access, or multi-step reasoning, since each dispatch then runs a full agentic loop.

Re-run with filter: { column: "<col>", exists: false } to reprocess only rows missing a result column. The run() result also returns deduplicated failure groups with error messages and affected row ids for inspection.

Batch sizes are clamped to 1-50, concurrency is capped at 10 dispatches, and console output is limited to about 5 KB. Tables are evicted after five tables per session, and subagents only see the instruction and context, never the agent's conversation.