Scry turns internet research into a bounded SQL workload for agents

Scry has launched an agent-facing search service that treats the public internet less like a list of links and more like a database. Its first-party launch page reports roughly 141.6 billion queryable rows across 43 sources at midnight UTC on Sept. 18, including Reddit, Hacker News, Stack Exchange, web crawls, books, papers and software repositories. Agents connect through MCP or an HTTP API and issue read-only SQL, typed search, vector queries and bounded Datalog-style programs against source-native relations (Scry launch page; live corpus census).

That makes Scry relevant to data-agent teams even though it is not a conventional enterprise NL2SQL product. The model is still responsible for reading a schema and composing the query, but the target is a shared internet index rather than a customer warehouse. The service’s stated contract requires a row limit and places each statement under deadline, memory and budget controls. Results retain the identifiers and provenance fields exposed by their source relation, giving an agent a path back to the underlying record (Scry launch page).

Search becomes a compute job

The unusual part is not merely SQL access. Scry documents recursive SQL plus fixpoint graph programs for traversing citations, replies, quotes and other edges. Programs declare a depth cap of one to 64 rounds, while request headers can set time and spending ceilings. The response reports which bounds truncated a run. That turns a potentially open-ended research request into an inspectable compute job: operators can see the query, corpus, limits and whether execution completed within them (recursive-search documentation).

The practical distinction from ordinary web search is aggregation. An agent can ask for counts, intersections, grouped trends or bounded graph walks over all matching records rather than repeatedly fetching top-ranked pages. Scry’s live census separately reports source-specific row counts, observed extents and known holes; that matters because a zero in the index is not necessarily proof that nothing exists at the original source (live corpus census).

Promising evidence, with an alpha-sized caveat

Scry also publishes a self-run result on Google’s 900-question DeepSearchQA benchmark. It reports 71.8% fully correct for a GPT-6 Astra agent using Scry, versus the 66.1% Gemini Deep Research Agent result shown on Google’s Kaggle leaderboard. The vendor exposes all 900 grades, its frozen configuration and tool-call transcripts, but explicitly notes that the benchmark has no official submission process. The comparison should therefore be treated as reproducible vendor evidence, not an independently certified leaderboard result (Scry benchmark page).

The service is also explicitly alpha. Its terms make no availability commitment, say the corpus and interfaces may change, and prohibit bulk reconstruction or redistribution of a substantial part of the indexed corpora. Paid use draws from prepaid credits and may include congestion pricing (Scry terms).

For practitioners, the launch is a useful architectural signal: agent search can be governed with many of the same primitives as database access—schemas, read-only execution, query plans, resource ceilings and provenance. The remaining question is whether Scry’s rapidly changing corpus and alpha service can deliver the predictable coverage and availability that production research agents require.