| name | web-research |
| description | Look something up on the web and present a sourced answer, optionally opening the top result in a spatial browser. Use for factual questions, definitions, how-to lookups, comparisons, and "search for…/look up…/find information on…". Triggers: search, look up, find info, what is, who is, how do I, research, google, current weather/facts. |
| license | MIT |
| compatibility | Requires JarvisVR agent-backend. Offline/MockLLM returns deterministic results; set a real provider/key to fetch live data (see system-agent configure-llm). |
| metadata | {"agent":"research-agent","category":"research","version":"1.0","author":"jarvisvr"} |
| allowed-tools | web_search show_web show_panel |
Web Research
Turn a question into a short, sourced spoken answer plus an optional artifact the
user can open and keep.
Steps
- Form a tight query. Extract entities and intent; drop filler. For "current
X" questions (weather, scores, prices) prefer the dedicated specialist
(
research-agent's market-briefing, or web-research for general facts).
- Call
web_search{query}. Returns data.results (title, snippet, url) and
a web_panel directive for the top hit.
- Synthesize a 1–2 sentence answer in
agent.speech; cite the source name.
- Surface an artifact: keep the
web_panel for browsing, or summarize the
findings onto a panel (show_panel) when reading beats browsing.
- If the user wants depth on one result, hand to
summarize-source.
Output
agent.speech:
{ "text": "Mixed reality blends real and virtual so they coexist and interact —
here's the overview from Wikipedia.", "final": true }
web_panel (show_web / from web_search):
{ "widget_type": "web_panel",
"props": { "url": "https://en.wikipedia.org/wiki/Mixed_reality", "title": "Search: mixed reality",
"interactive": true } }
Summary alternative (show_panel → panel):
{ "widget_type": "panel",
"props": { "title": "Mixed reality", "body": "Definition, history, and devices.",
"sections": [ { "heading": "Top sources", "text": "Wikipedia, Nature, BBC" } ] } }
Edge cases
- No/low-quality results → say what you tried and offer to refine the query;
don't fabricate facts or URLs.
- Time-sensitive ("today", "latest") → note recency limits when offline; live
fetch requires a configured provider.
- Multi-source synthesis (compare 3 articles) → delegate to
summarizer
sub-agents (one per source) and merge — see orchestration agent-routing.
- News specifically → use
news-digest; markets → market-briefing.
- Reading vs browsing → long answers belong on a
panel/document_viewer,
not read aloud in full.
