subagent-delegation

name	subagent-delegation
description	Decide when and how to delegate work to specialist subagents and model-profile subagents.

Subagent Delegation

The orchestrator can hand off specialized work via task(subagent_name, instruction). Use delegation when a task has a clear specialist owner or when parallel work would improve latency/quality.

Specialist subagents

• researcher — multi-hop web and document research; returns cited notes.
• analyst — quantitative analysis, ratios, and data-frame operations.
• charter — plotting, chart design, and visualization artifacts.
• pdf_reader — deep reading and extraction from uploaded PDFs.

Model-profile subagents

• deepseek-v4-flash — fast multilingual reasoning.
• deepseek-v4-pro — high-accuracy reasoning and coding.
• nemotron-3-super — agentic reasoning and tool use.
• nemotron-ultra — largest Nemotron model for complex analysis and coding.
• nemotron-3-ultra — frontier long-context reasoning (Nemotron-3 Ultra 550B).
• nemotron-3-nano — low-latency text tasks, summaries, quick answers.
• mistral-small-4 — multimodal reasoning over images + text.
• llama-4-maverick — vision and long-context understanding.
• gemma-4 — lightweight multimodal OCR and visual Q&A.
• gpt-oss-120b — tool-heavy structured output tasks.
• glm-5.1 — long-context text synthesis.
• qwen3.5 — coding and SQL generation.
• minimax-m3 — multimodal vision-language tasks.
• step-3.7-flash — fast multimodal vision-language work with native tool use.

Delegation rules

1. One specialist per distinct sub-task; do not nest more than two levels.
2. Include all context the subagent needs in the instruction (widget IDs, PDF names, question, desired output format).
3. If a subagent fails, switch to a different profile or do the work yourself instead of retrying identically.
4. Always integrate the subagent result into the final answer and cite sources.