LLM Modes and Modalities

Definition

Modes (e.g. chat, “agent,” “deep research,” coding assistants) are packaged behaviors: prompt templates, allowed tools, retrieval sources, timeouts, and sometimes a different base model. Modalities include text, images, audio, and video: image and video generation typically use diffusion or video models distinct from the text LLM, orchestrated by the same product shell.

Why it matters

Buyers compare brands; operators must compare capabilities and limits. A “deep research” preset may issue many tool calls and burn tokens; “agent mode” may widen tool scope and increase risk if approvals are weak.

When to use

Use research-style modes when you need multi-source synthesis with explicit steps and citations. Use agent-style modes when tasks require iterative tool use. Use image/video modalities when the deliverable is pixels or clips, not prose—still subject to policy and copyright norms.

When not to use

Do not enable broad agent modes for production side effects without scoped credentials and logging. Do not assume one model does text and generation equally; latency and cost profiles differ.

Failure modes

Treating marketing mode names as architecture; unbounded browse-and-summarize loops; generated media shipped without provenance or human review where required.

Related pages

Choosing LLM tiers and families · Research synthesis case study · LLMs in agentic systems · Categories