Search is not evidence: a routing model for CLI research agents
A research agent can return a convincing answer while never reaching the source page that actually supports it. Search results are discovery, not evidence.
The workflow we are testing with Gemini CLI uses a routing ladder for every source:
- Prefer an official API or direct fetch.
- Use protected-page extraction when ordinary retrieval is blocked.
- Open a real browser only when JavaScript, filters, login, or interaction requires it.
- Hand control to a person for authentication, consent, CAPTCHA, or other trust boundaries.
Every worker returns the same record: URL, retrieval time, page state, evidence, recovery history, and a blocked status when the boundary cannot be crossed safely. The coordinator synthesizes records rather than uncited model memory.
The full Gemini CLI + BrowserAct workflow is here: https://www.browseract.com/blog/gemini-cli-browser-automation
What evidence schema do you use for agent-generated research?
Trending on Indie Hackers
the routing ladder is the right instinct. the catch on the evidence side is that a URL isn't evidence either. an agent can fetch the right page and still assert something the page doesn't literally say, or the page changes later and the citation quietly rots. what saved us building a local-first research assistant was binding every synthesized claim to a specific quoted span plus a content hash of what we fetched, not just the source url. verification then becomes a diff against that snapshot instead of a fresh search, and a stale cite surfaces as a hash mismatch rather than a confident wrong answer. so the schema is claim to span to hash, and the coordinator won't assert anything it can't point at.