Research current public sources¶
examples/public-web/humagen.yaml gives a native agent, researcher, the
web_search and web_fetch tools. It searches current public sources, opens
the pages it relies on, and reports the supporting URL beside each claim. The
URLs are answer text the model wrote, not platform-verified citations.
Both tools run on the Humagen server, not in a sandbox, and need no account
or search key. web_search returns up to 10 results, each a title, an address
and a snippet, from DuckDuckGo's no-script page, which may refuse an automated
client; the tool then says so. web_fetch reads one public http or https page as text, at
most 40,000 characters; it refuses addresses inside a private network and
content that is not text, such as a PDF or an image. Every result is marked as
web content, not instructions.
Before #383 this example ran on the Claude Code program's own WebSearch and WebFetch tools.
Start the researcher¶
Set a provider key in the shell (ANTHROPIC_API_KEY, OPENAI_API_KEY or
GOOGLE_API_KEY; the agent tries them in that order), then:
humagen config check -f examples/public-web/humagen.yaml
humagen start -f examples/public-web/humagen.yaml
The profile uses development sign-in and memory state for trusted local trials: every client reaching its API is an administrator. For a shared service, put the agent in an authenticated workspace. Retrieved text is evidence, not authority to change the task.
Ask for traceable evidence¶
Name the question and the authoritative sites. Ask the researcher to search, open the pages it relies on, and put a supporting URL beside each factual claim. A search snippet alone does not establish what the current page says. Ask for an unknown when a field is not in the retrieved material.
For example, find the current stable Go release on go.dev, open its download
feed, and report the version with the exact opened URL. Ask for a release date
only when an opened source supplies it. Read the run's tool events to confirm
that web_search and web_fetch ran; a plausible answer alone is not proof.
Limits¶
A native turn makes at most 10 rounds of tool calls, so a question that needs many pages takes more than one turn. The researcher has no signed-in browser tab, JavaScript, screenshot or PDF; use the browser worker for those. A failed fetch is a failed retrieval, not evidence that a source confirms a claim.
The tools were checked against the live web on their own
(HUMAGEN_WEB_LIVE=1 go test ./pkg/webtools -run TestLive). This profile has
not been run with a live model; the earlier end-to-end check ran on the removed
Claude Code path.