Edison Scientific (the commercial spinout of FutureHouse) sells “AI agents that do scientific research.” Their flagship, Kosmos, claims a single run approximates months of PhD work — at $200 per run. But the platform also exposes a lighter agent called Literature, which answers scientific questions with fully cited responses, and academics get free credits.

I’m a master’s student trying to get into data systems and data security. My bottleneck is field knowledge: I don’t know what the open problems are, which papers matter, or where the gaps are. So I ran an experiment — could this thing compress the “survey a new field” step that usually takes weeks of skimming Google Scholar?

The setup

One API call through their Python client (edison-client, Apache 2.0):

from edison_client import EdisonClient, JobNames, TaskRequest

client = EdisonClient(api_key="...")
task = TaskRequest(
    name=JobNames.LITERATURE,
    query="What are the main open problems in data lineage "
          "tracking for scientific datasets? Give a brief overview.",
)
response = client.run_tasks_until_done(task)

The task ran for a few minutes — it was actually searching a corpus of ~175M papers, not autocomplete-ing an answer.

What came back

A structured survey of eight open problems in data lineage, each with citations to specific pages of real papers:

1. End-to-end completeness — manual work, wet-lab steps, and ad-hoc scripts escape capture; retrospective reconstruction from logs is inherently incomplete
2. Granularity vs. cost — fine-grained lineage explains more but can produce provenance larger than the data itself
3. Interoperability — W3C PROV exists but adoption is inconsistent across domains
4. Federation — lineage crosses instruments, clouds, and institutions; queries over distributed provenance are hard
5. Versioning — code, containers, and infra evolve; recording execution alone doesn’t guarantee reproducibility
6. Scalable querying — big lineage graphs are expensive to store and cognitively hard to read
7. Quality and trust — captured lineage can itself be wrong; there are no standard metrics for completeness or authenticity
8. Privacy vs. traceability — detailed lineage can expose sensitive data; access control and selective disclosure remain open

Every claim carried a citation — real papers with DOIs, citation counts, and journal information (Johns et al. 2023 in JMIR, Khan et al. 2019 in GigaScience with 109 citations, Souza et al. 2022 on ML workflow provenance).

Honest critique

• The citations are real and page-level. I checked a few DOIs — they resolve. This is the feature that separates it from a chatbot.
• It’s a starting map, not a verdict. One cited source was a 2026 preprint with an “unknown journal” tag — the agent includes fresh but unvetted material. You still have to read the papers.
• Depth has a ceiling. It synthesizes what reviews already concluded; it won’t notice a problem nobody has written about yet.
• Cost model matters. Literature queries are cheap-ish and free-tier-able. Kosmos at $200/run is a different budget category.

What this changes for me

The useful thing isn’t the summary — it’s the reading list with a structure. Each of those eight problems is a direction I could pick for a project, and now I know which 11 papers to read first. The step “what should I even look at?” went from weeks to one afternoon.

It also quietly validates the path I’m on: half the open problems it listed (federation, privacy-aware queries, versioning) are systems problems, not science-domain problems. The people who will solve them are data infrastructure engineers — which is exactly the job family I’m aiming at.

If you’re entering a new field, this class of tool is worth an hour of your time. Not because it does the reading for you — it doesn’t — but because it converts “I don’t know what I don’t know” into a checklist.

Tools used: Edison platform (platform.edisonscientific.com), edison-client Python SDK. Free academic tier.