Key Features

• Asynchronous: Submit instructions via POST /research/v1 and receive a researchId immediately. Poll or stream for results — tasks can run from seconds to minutes depending on complexity.
• Multi-step reasoning: The agent iteratively plans, searches, crawls, and synthesizes. It decomposes your instructions into sub-tasks, runs them in parallel, and combines findings into a single output.
• Structured output: Optionally provide an outputSchema (JSON Schema) to get validated, parsed JSON back instead of free-form text.
• Streaming: Pass ?stream=true when polling to receive real-time Server-Sent Events as the agent thinks, searches, and writes.

How It Works

1. Create a task with your instructions and optional output schema via POST /research/v1.
2. The agent plans by decomposing your instructions into targeted sub-tasks.
3. Each sub-task searches the web and crawls relevant pages to gather information.
4. Findings are synthesized into a final output with citations.
5. Retrieve the completed result via GET /research/v1/{researchId} or list all tasks with GET /research/v1.

Models

Task Lifecycle

Tasks progress through these statuses: pending → running → completed | failed | canceled. Only completed tasks are billed. The costDollars field on completed tasks provides a detailed breakdown of searches, pages crawled, and reasoning tokens used.

Next Steps

• Create a research task
• Get a task by ID
• List all tasks