Agent Science

Inspiration

A citation can make an answer look settled long after its evidence has changed. Repeating the search loses the useful part of the previous investigation: the exact source, what it established, and what remained uncertain.

Agent Science is a websearch truth companion. For this entry, the concrete use case is factual production: checking claims before they reach a documentary script or another published story.

What it does

The reviewed implementation retains claims with source excerpts, retrieval timestamps and document hashes. A later recheck distinguishes unchanged documents, changed documents whose cited span remains, and cited spans that have disappeared. Missing evidence moves a result to UNKNOWN; it does not establish that the claim is false.

The method remains inspectable. A source's hostname is a classification hint, not proof of the speaker's identity or the truth of every statement it contains.

How we built it

The Python clearance pipeline connects source discovery through Parallel with retained evidence. A Google ADK wrapper uses Gemini to call the script-clearance tool and returns the tool's structured report. It rejects a model response that never produced the required tool result.

Cloud Run hosts a separate private-workspace application backed by Cloud Storage. Workspace research uses isolated request caches and bounded research budgets. Connecting the submitted hosted journey to the required Gemini/ADK and Parallel runtime is an outstanding release check; the existence of the wrapper is not proof that every hosted route invokes it.

Challenges

The cold release review found that the public judge URLs led to login and that the source classifier could treat example.gov.attacker.invalid as a primary source. The reviewed candidate uses parsed hostnames and domain boundaries, with controls that fail against the old implementation.

The same review separated four states that had been blurred: code tested locally, code published for review, the revision actually running, and a complete competition entry.

What we learned

Reuse needs a clock and a way to revisit evidence. Finding the same sentence again establishes that the sentence is present, not that it is true. Multiple links do not by themselves establish independent corroboration.

Current release status — September 8

The public Cloud Run journey now supports asking a fixture question, inspecting its exact passage and trust card, saving an answer, changing a synthetic source and observing the previously held answer become unsettled. The new 2:06 video shows this hosted flow using actual browser screenshots and locally generated narration.

Try it at https://agent-science-33kamss2jq-uc.a.run.app/journey. This public flow uses labelled synthetic documents and shared, ephemeral demo state. It makes no live Gemini or Parallel call. It demonstrates evidence retention and correction behaviour, not measured search savings or independent truth verification.

The public repository is MIT-licensed. Its default branch remains ea05a10; PR 2 remains open. The hosted revision and default source branch still need reconciliation, and the current hosted live-provider journey remains a release check. Historical partner-run reports are not presented as a live run of this public demonstration.

What's next

Make the media-review journey work end to end on the hosted release, then extend source freshness into a reusable collection of claims with explicit uncertainty and correction history.

Built With

• gemini
• google-adk
• google-cloud
• google-cloud-run
• parallel
• python
• sqlite