The missing control plane in web research agents
Why a research agent should separate discovery, evidence retrieval, archive scope, and media processing instead of hiding everything behind one crawler.
Most research agents can reach the web. The harder problem is knowing when a result is only a lead, when a page has actually been read, and when a download or archive action is allowed. yichen-web-research addresses that problem as a control plane rather than another universal crawler.
Research has distinct risk levels
Public discovery, reading a known link, exporting private bookmarks, and transcribing local media are different operations. The repository routes them to separate skills instead of hiding them behind one permissive command. That makes scope visible to both the operator and the agent.
Evidence should survive the handoff
The search child produces candidates with URLs, provenance, backend information, and limitations. A later archive step receives only supplied or confirmed targets. This prevents snippets from being promoted into facts and gives a report a traceable path from question to source.
For a technical investigation, this means a repository README, release note, issue, or official documentation page can remain distinct evidence objects. A blocked page, timeout, rate limit, or private target should remain an explicit failure state instead of being silently filled by model knowledge.
Platform rules stay local
X search, known X URL reading, public WeChat access, and browser-session search each have different routes and authorization requirements. The project keeps those differences rather than flattening them into a generic crawler. No CAPTCHA, login wall, rate limit, or private boundary is bypassed.
Tests define the boundary
The doctor and validation scripts check that the family is complete and that search, archive, bookmark export, and ASR do not silently cross into one another. This makes a safety rule regressible instead of aspirational.
Operational conclusion
Use the top-level router for multi-stage research. Use the child skill directly for a single action. The durable lesson is simple: a research agent becomes more trustworthy when it records what it found, what it actually read, what it could not access, and what it was not authorized to do.