When an agent needs to answer a research question, it does not need a list of documents. It needs an answer — one it can trust, cite, and build on.
But the tools we give agents today are built for humans. Web search, Tavily, Exa, Brave — all return document lists. The agent then spends tokens parsing HTML, reconciling contradictions, guessing at provenance, and connecting the dots. Every retry costs money. Every quietly wrong answer costs trust.
That is a structural tax on every agentic workflow. And it is invisible — no one notices the cost because there is no alternative to compare against.
What a library looks like
Alexandria inverts the model. Instead of returning documents and letting the agent extract knowledge, we archive knowledge directly as structured entities and relations — typed, versioned, provenance-backed, confidence-rated.
- A Paper node stores title, authors, date, summary, and direct citations. It does not hide inside an HTML page with a JavaScript paywall.
- A Strategy node stores the paper that introduced it, the dataset it was tested on, and the results it produced — all as typed edges, not flowing prose.
- A Result node stores a quantitative finding with exactly one number, one confidence rating, and one citation. No ambiguity about what was measured.
- A Belief Difference edge flags when two sources contradict each other. The agent sees the conflict explicitly instead of having to infer it.
Each entry is validated against a schema. Aliases are resolved. Confidence is rated (never inflated). Provenance is mandatory. The result is not a document — it is a machine-readable fact.
One call, one answer
The library's anchor tool, library_assemble, takes a target string and returns one compact, cited answer bundle. It resolves the entity, walks the typed graph one hop deep, dedupes by alias, groups neighbors by role (support, risks, assumptions, depends_on), and surfaces belief differences. The agent reads once.
A token estimate is included so the agent can decide whether the answer is worth reading before it reads it.
Capabilities no search tool has
Because the data is a graph, the library can answer questions no web search can:
- Pathfind: "How is momentum investing connected to CAPM?" — the library returns the shortest typed path through the graph (momentum → dataset → Fama-French → CAPM), with every hop labeled. No search engine can do this because no search engine has a typed graph.
- Compare: "What is the relationship between Jegadeesh-Titman 1993 and Moskowitz-Ooi-Pedersen 2012?" — the library returns direct edges, shared neighbors, belief differences, and provenance overlap in one call.
- Trending: "What are other agents asking about right now?" — the library returns the demand map from its own telemetry: top types, tags, routes, and query volume.
- Traverse: "What does the neighborhood of momentum investing look like?" — the agent walks the graph interactively, one hop at a time, with typed relation labels.
These are not search features. They are graph operations. No MCP-first competitor (Tavily, Brave, Exa, Perplexity, Search1API) has any of them.
How agents pay
Every call is chargeable. The agent attaches a proof of payment (dev token in testing, stablecoin authorization in production), or receives a 402 Payment Required response with the exact price and where to pay. No signup, no API key, no human in the loop. Payment is the credential.
The library is live in seed form with 334 nodes and 440 edges across quantitative finance and machine learning. The same pipeline scales to any shelf. Agents query it over MCP or our HTTP API.