deep-research

name	deep-research
description	Queue deep, multi-step research and analysis tasks to the DeepAgents worker. Supports execution depth presets, request-specific rubrics, SubAgent delegation, and RubricMiddleware cross-validation.

Deep Research

Delegate complex, long-running research and multi-document analysis to the DeepAgents background worker service.

Routing Rule

For deep research, comprehensive analysis, or multi-step investigations, execute /sandbox/bin/deep-research.

Do not perform the same long-running task again inline after it has been delegated.

How To Use

/sandbox/bin/deep-research [--depth shallow|standard|deep] [--rubric "<custom criteria>"] "<research prompt or goal>"

Examples

Standard Research (Default):

/sandbox/bin/deep-research "Analysis of vector database performance benchmarks in 2026"

Deep Research:

/sandbox/bin/deep-research --depth deep "Research the 5 levels of agentic workflow platform security and draft a technical whitepaper"

Custom Rubric Validation:

/sandbox/bin/deep-research --depth deep --rubric "Compare the alternatives in a table and cite each major factual claim" "Compare public retrieval architectures"

Options & Arguments

• --depth <shallow|standard|deep>: Execution depth preset:
• shallow: 25 graph steps, 1 rubric iteration
• standard (default): 50 graph steps, 2 rubric iterations
• deep: 100 graph steps, 3 rubric iterations
• --rubric "<text>": Optional custom quality rubric for RubricMiddleware cross-validation. If omitted, the worker creates a request-specific evidence rubric before invocation.
• Position 1: The research prompt or goal (required string)

How It Works

1. Task Enqueued: Submitted to the DeepAgents worker service via SQLite queue.
2. Planning: The worker initializes write_todos and applies either the supplied rubric or a request-specific default rubric.
3. SubAgent Delegation: Research subtopics are delegated to isolated SubAgent workers to prevent context window saturation.
4. Cross-Validation: RubricMiddleware evaluates the output against the quality rubric, triggering iterative refinement loops if criteria are not met.
5. Output Delivered: Formatted report is returned to stdout.

Execution Constraints

• Up to 5 worker threads run in parallel (DEEPAGENTS_WORKER_CONCURRENCY=5).
• Task results are retained for 7 days (DEEPAGENTS_TASK_TTL_HOURS=168).
• Only the built-in read-only web-search and document-search tools are available.