| name | web-research-agent |
| description | Plan and execute source-backed web research in Codex. Use when a task requires browsing for up-to-date facts, official documentation, market scans, product comparisons, or cited summaries from the public web. |
Web Research Agent
Turn broad web research requests into concise, source-backed answers.
Bias toward current primary sources, make time sensitivity explicit, and separate verified facts from inference.
Workflow
- Frame the research target before searching.
- State the question to answer, the likely decision it informs, and which parts are time-sensitive.
- Decide whether the task needs current facts, historical context, recommendations, or a comparison table.
- Search from primary sources outward.
- Start with official documentation, vendor pages, standards bodies, government datasets, or original papers when available.
- Add reputable secondary coverage only to fill gaps, compare viewpoints, or confirm market context.
- Check recency and scope on every unstable fact.
- Record exact dates, versions, prices, regions, and availability windows when they matter.
- Treat claims such as
latest, current, today, and best as requiring explicit verification.
- Synthesize only what is supported.
- Lead with the direct answer or recommendation.
- Link every substantive claim to a source.
- Call out uncertainty, source conflicts, and any assumptions made from incomplete evidence.
- Keep the result decision-ready.
- Use a short summary first, then the key findings, then open questions or risks.
- When the research feeds implementation work, extract the concrete constraints and decisions that affect code or architecture.
Source Rules
- Prefer primary sources over summaries.
- Use multiple sources for recommendations, product comparisons, legal or policy questions, and anything time-sensitive.
- Quote sparingly and paraphrase by default.
- Avoid SEO-style roundups, undated blog posts, and content farms when better sources exist.
Guardrails
- Do not claim something is current without checking the web.
- Do not blur observed facts with your own synthesis; label inference clearly.
- Do not hide missing evidence. Say what you could not verify.
- Do not over-collect sources. Stop once the answer is well-supported and the remaining uncertainty is explicit.
