iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
To give an AI agent current web information without flooding its context window, enable a hosted web-search tool and control how much retrieved content reaches the model. OpenAI documents a search context-size setting; Anthropic documents dynamic filtering that can discard irrelevant results before they enter context. Neither provider documents a guaranteed token-savings figure for this workflow, so measure your own agent rather than promising a percentage.
How live web search works in an agent
Live search is a tool integration, not an instruction you can switch on by wording a prompt. The agent needs access to a supported search tool in its API request. For example, OpenAI’s Agents API web-search documentation says omitting the web_search tool turns built-in search off, and shows live mode as a configured tool option.
When enabled, search can retrieve current web information for a question or task. Anthropic describes its web-search tool as providing Claude access to real-time web content beyond its knowledge cutoff in its Claude API web-search documentation. Actual availability depends on the current API, supported models, and tool versions described by each provider.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Fix the driver behind crashes, sound loss and screen glitches3Clear out junk files and repair common Windows errorsControl how much retrieved content reaches context
Set a context-size limit
OpenAI documents context_size choices of low, medium, and high; medium is the default in the cited documentation. A lower setting is a sensible starting point for straightforward questions, while a task that needs broader evidence may require more context. The setting controls search context, but the documentation does not quantify token savings for a particular workload.
#1 Best Overall
Filter results before they enter the model context
Anthropic documents dynamic filtering for a newer version of its search tool. Code can retain relevant results and discard irrelevant material before it reaches the context window. This can help when searches return more material than an answer needs; confirm that the relevant model and tool version support the feature in the current web-search documentation.
Use search first, then fetch only when needed
Search and page fetching solve different problems. Search helps locate relevant pages and surface evidence; fetch retrieves content from a specific page when the search output is not enough. Anthropic documents web fetching and filtering fetched content in its web-fetch documentation.
- Enable the supported search tool. Configure it in the agent’s API request and verify that the request actually includes the tool; a prompt alone does not enable it.
- Start with limited context. Use an appropriately conservative context-size setting or filtering strategy for the task.
- Inspect the search evidence. If it adequately answers the question, avoid fetching whole pages unnecessarily.
- Fetch a specific page if needed. When the search result lacks sufficient evidence, retrieve the relevant page and use filtering if the fetch tool supports it.
- Keep source citations. Preserve citations and check that the cited evidence supports the answer. Both providers describe citations in their web-search documentation.
Choose a setup by its controls, not a promised savings figure
The provider documentation describes different controls, so compare the features relevant to your agent rather than assuming one configuration is universally best.
Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchPC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11| Capability | OpenAI documentation | Anthropic documentation |
|---|---|---|
| Search access | Live, cached, and disabled modes are documented. | Real-time web search is documented; live-versus-cached modes are not stated in the cited web-search documentation. |
| Context control | context_size supports low, medium, and high; medium is the documented default. |
Dynamic filtering is documented for a newer search tool version. |
| Fetch specific pages | Not stated in the cited web-search documentation. | A separate web-fetch tool is documented. |
| Citations | Citations are described in the web-search documentation. | Citations are described in the web-search documentation. |
| Model and tool support | Check the current documentation for supported API behavior and parameters. | Support depends on the documented model and tool version; check the current documentation. |
Measure token use and answer quality on your workload
Reducing unnecessary retrieved context is a mechanism for using fewer tokens, not proof of a particular saving or a better answer. The cited provider documentation does not publish a token-savings statistic for this exact workflow. Evaluate representative tasks before setting expectations.
- Record input and output tokens for each run.
- Check whether citations are relevant and support the answer.
- Assess whether the answer is complete for the task, not merely shorter.
- Track latency alongside token use, since extra search or fetching can affect response time.
Compare runs using the same representative queries and task criteria. Increase context or fetch a page when the evidence is insufficient; keep tighter limits when the answer remains complete and well-supported. Tool versions, supported models, parameters, defaults, pricing, and operational constraints can change, so consult the linked official documentation for current details.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
