In short
Save agents and deep research modes for tasks that genuinely need many steps, such as researching a broad topic across many sources, and use an ordinary chat for quick questions. Agents call a model many times per task: Anthropic reported agents using about four times the tokens of a chat, and multi-agent systems about fifteen times.
Steps
- Try an ordinary chat or a search first, and use an agent only if that falls short.
- Give the agent a clear goal, a scope and a point at which to stop.
- Tell it which sources or sites to use, so it doesn't wander.
- Watch the first run and stop it if it starts repeating itself.
- Reuse the report it produces instead of re-running the same research.
Agents are powerful but expensive, so brief them well and use them for big jobs. An AI agent, including the "deep research" modes in many chatbots, plans, searches, reads, writes and checks in a loop. Each step is another call to a model, usually carrying the growing history of the task.
How much more do agents use?
Many times a chat. Anthropic, describing its own multi-agent research system in 2025, reported that agents typically used about four times more tokens than chat interactions, and multi-agent systems about fifteen times more. Kim, Shin, Chung and Rhu found that agentic workflows used substantially more energy than single answers, with diminishing returns in accuracy as more computation was spent.
That is a fair trade for a task that would take you hours. It is waste for something a single question would answer.
How do I brief an agent well?
- Say what done looks like: "A one-page summary of the five most cited studies since 2023."
- Set the scope: time period, countries, types of source.
- Point to sources you trust, if you know them.
- Ask for a short output. A long report costs more to write and to read.
Common mistakes
- Using deep research for simple questions that a search or chat handles in seconds.
- Re-running the same research instead of saving and updating the first report.
- Letting a stuck agent carry on. If it keeps revisiting the same pages, stop it and rephrase.
If you build agents, see stop AI agents wasting tokens. For the wider issue, read agentic AI and token waste.
Sources
- How we built our multi-agent research system, Anthropic (2025)
- The cost of dynamic reasoning: demystifying AI agents and test-time scaling from an AI infrastructure perspective, Kim, Shin, Chung and Rhu, HPCA 2026 (2025)
Related
How do I know when to use an AI "thinking" or reasoning mode?
When an AI reasoning or 'thinking' mode is worth it, when a standard model will do, and why reasoning can use many times more energy per answer.
How do I stop AI agents wasting tokens?
Practical guardrails for AI agents: stop conditions, step and token limits, loop detection, context trimming and small models for simple steps.
Agentic AI and token waste
Why AI agents can use many times the tokens and energy of a chat, how loops waste them, and the guardrails that keep agents right-sized.