Edison's Kosmos: 57.9% of Interpretations Held Up

Key points

• A researcher enters a question, and agents run the literature search and the data analysis.
• Every sentence in a report links straight to the notebook code or paper behind it.
• Human review focuses on interpretation of results, where reliability is lowest at 57.9%.

Agents take over the middle of research, which eats the most lab time

To test one hypothesis, a lab has to read a great many papers and write analysis code. It also has to clean up experiment data, run statistics software, and draw graphs. In university labs and pharmaceutical R&D groups, graduate students, postdocs, and staff scientists have done this middle stage by hand.

Edison Scientific (Edison, for short) wants to hand this middle part of research to agents. Edison is a for-profit spinout that was set up in 2025 to sell the research agents of FutureHouse. FutureHouse is a nonprofit AI research lab that Sam Rodriques and Andrew White co-founded in 2023. Sam Rodriques is now Edison’s chief executive officer (CEO). According to a December 18, 2025 report by the tech outlet SiliconANGLE, Edison raised a $70 million seed round led by Triatomic Capital and Spark Capital.

Edison’s core product is Kosmos, which the company calls an “AI Scientist.” Kosmos first launched on November 5, 2025, as an improved version of a system FutureHouse built earlier. Before that, an earlier system called Robin had shown the same way of working on a single drug-discovery project. Robin was published in the journal Nature on May 19, 2026.

Besides Kosmos, Edison also offers smaller agents, each dedicated to literature search, data analysis, checking a study’s novelty, or chemistry analysis. Since May 2026, it has been widening its scope to drug-development work in general, such as toxicology assessment, writing clinical trial protocols, and writing documents for regulators.

Enter one objective, and it runs tens of thousands of lines of code and returns a cited report

In the early version of Kosmos, a researcher started the work by specifying a research objective and a dataset. Given the objective and data, Kosmos repeated cycles for up to 12 hours. In each cycle, it ran up to 10 literature search and data analysis tasks at the same time.

The agents summarized their results and recorded them in a shared knowledge space called the “world model.” The world model is an internal store that lets the data analysis agents and the literature search agents share what they have found. The paper does not describe how it is built inside. Kosmos consulted the world model, asked itself what to do next, and carried on to the next cycle.

By the time Kosmos finished, the scale of what it had run was large. On average, one run used about 166 data analysis rollouts, which ran about 42,000 lines of code. It also used 36 literature search rollouts, which read about 1,500 papers. When everything was done, Kosmos wrote three or four research reports. Every sentence in a report carried a citation, either to the academic paper behind it or to the Jupyter notebook file where the actual calculation was recorded. A reviewer could open the code or the source text that supports the sentence right away.

In this early design, the human scientist’s role centered on the start and the end of the work. Before the work began, the scientist had to write the research objective and prepare the data. According to Edison’s guide, the objective should be something a collaborator would need weeks or months to finish, and it should be written as specifically as if explaining it to an experienced fellow scientist. The data for analysis must be under 5 GB, cleaned of errors in advance and clearly labeled. The guide carries this warning: “We do not recommend getting Kosmos to process your data.”

During the run itself, people could do nothing. The launch paper also explains that the system was designed so scientists cannot interact with Kosmos in the intermediate cycles. The scientist’s role started again only after the results came out. In the paper’s words, “the scientist’s role shifts to evaluation and interpretation of results.” Edison’s team sometimes ran the same objective several times to see which different paths Kosmos explored each time. A launch-time blog post compared Kosmos less to an everyday chat tool and more to a kind of lab reagent kit, which you bring in when an important research objective comes up.

How the work divides once real lab work is added was clear in the case of the earlier system, Robin. Robin linked three agents, one for literature review, one for judging candidate molecules, and one for analyzing experimental data. It was put to work on finding a treatment for dry age-related macular degeneration. Robin led the process of proposing a biological mechanism, choosing the drugs to test, analyzing the experiment data, and suggesting the next experiment. FutureHouse said: “Human researchers executed the physical experiments, but the intellectual framework was entirely AI-driven.” The researchers explained that Robin handled everything from forming hypotheses and choosing experiments to analyzing data and making the figures in the main text of the paper. They said it took only 2.5 months from conceiving Robin to submitting the paper to a journal.

A lone agent was rebuilt to talk like a fellow researcher

In May 2026, Edison overhauled how Kosmos works. In a technical blog post, the developers looked back on the old version: “You put in one prompt and it worked for days.” But this kind of deep analysis wasted too many resources on light questions.

The rebuilt Kosmos became “a persistent code-writing LLM agent that collaborates during the research process.” Kosmos now first discusses the research approach with the researcher, and it takes feedback mid-work and changes direction. While the work runs, it builds a web page that shows research progress, and it emails the researcher when relevant new papers are published. It connects to Slack, the messenger program, so it can work with a whole research group. Work that has to run in parallel, such as literature search, is split among subagents in separate isolated environments.

Features for companies were added too. It can now connect to a company lab’s electronic lab notebooks and assay data from biological samples, and customers can run Kosmos on their own servers. It can also be further trained on the customer’s own data.

The role of people is changing in the paperwork of drug development too. According to a case study Edison published on September 30, 2026, Population Health Partners, a firm that builds drug companies, brought in Kosmos to write the regulatory documents for a program entering the clinic. The two companies first built the scoring benchmark for this writing together, and Edison trained Kosmos on it. After that, Kosmos drafted key documents such as the nonclinical overview and the investigator’s brochure. Edison said that Kosmos wrote about 70% of the Pre-IND meeting request, a request for a meeting before filing an investigational new drug application with the US Food and Drug Administration (FDA), and the firm’s team edited the rest. In this process, the expert staff could spend their time cross-checking claims and figures and making final decisions, not writing drafts from scratch.

The analysis is accurate, but nearly half the sentences interpreting results missed

The biggest difference between a conventional lab and Edison is where people stand in the work. In an ordinary lab, the principal investigator asks a question, and junior researchers read papers, write code, run analyses, and report the results at regular meetings. Every step passes through human hands, and the principal investigator sees the combined results only after countless small decisions have already been made. Under Kosmos, agents handle those middle steps in parallel. The researcher asks the question at the very start and judges the results at the very end. In the latest version, the researcher also steers midway as a coordinator.

In a setup like this, checking the output becomes the core human job. Edison’s own evaluation shows why. Expert researchers reviewed 102 statements pulled from three Kosmos reports and judged whether each was supported or refuted by evidence. Overall, 79.4% of statements were supported.

But when the statements were split by type, the gap in reliability was clear. 85.5% of data analysis statements were supported, and 82.1% of literature statements were supported. By contrast, only 57.9% of statements that interpreted and synthesized results were supported. That means Kosmos does mechanical work such as calculation and citation well, but its accuracy drops when it explains what the results mean scientifically. The paper also admits that Kosmos “tends to make excessively strong claims,” and that it sometimes treats a result as a great discovery just because it is statistically significant, even though it is not important in real research. So people have to focus on catching errors of value judgment and interpretation, the point where the agent is weakest.

The citation system, which leaves a source for every sentence in a report, is an important safeguard that makes this kind of review possible. Every claim links to a Jupyter notebook or a published paper. So a reviewer can pull out just that claim and audit it, without redoing the whole study. Edison also checked whether results matched outside facts. Of the seven findings presented in the launch paper, three were cases where Kosmos reproduced, on its own, results from other teams’ unpublished data or preprints that it had never seen while running. In Robin’s case, the researchers ran wet-lab experiments with living cells themselves to test the agent’s hypotheses.

Outside scientists say this review work is still not easy. Georg Meisl of the University of Cambridge told the medical outlet Alzforum that for work that is hard to verify, “mistakes are still too common,” and that checking the output takes a lot of time. Lary Walker of Emory University offered a rule: “Use, but verify.” Manuel Lessi, on the biology community site preLights, said that on a research question whose answer he already knew, Kosmos’s results matched his own analysis. On a new question, though, checking the results took so long that doing the analysis himself might have been faster.

Edison priced research as a kind of consumable. At launch, Kosmos cost $200 per run, with free runs for academic researchers. Sam Rodriques told Alzforum that about 30,000 users in academia and biotech companies had tried Kosmos. The time estimates for the workload rest on calculations by the company and its collaborators. Beta users estimated that one run of 20 cycles equaled about six months of an individual researcher’s work. Edison’s own estimate, calculated from the number of papers read and analyses run, was about 4.1 months. After May 2026, Edison turned toward enterprise-focused contracts. It closed the first subscription sign-ups and gave academics one-time credits. It also formed a partnership with the pharmaceutical company Incyte, starting with target discovery and translational biology. The contract amount between the two companies has not been disclosed.

The standard for scarce skill could change, and paperwork could shrink

The skill most prized across a research organization could change. If agents can run literature search and code execution quickly in parallel, the value of the days a junior researcher spends cleaning data and writing analysis code will fall by comparison. What becomes more important is the eye of a senior scientist who can spot at a glance what is wrong among the plausible interpretations AI produces. The figure that only 57.9% of interpretation statements were supported suggests that this area of judgment will stay with people for the time being.

In pharmaceutical R&D and the service industries around it, document-heavy work could shrink. Writing reports for regulators in each country and reviewing clinical data consume huge amounts of cost and labor in drug R&D, and they are often handed to specialist writing teams or outside contractors. If the Population Health Partners case can be repeated in other research programs, the writing staff would spend more time on fact-checking and strategic decisions than on drafting. The collaboration model Edison set up with Incyte points to another implication. If a drug company further trains an agent on its own experiment data, the experiment data the company has built up over many years can become a lasting competitive advantage.

Other organizations that want to bring agents into research can take three lessons.

• Make every claim traceable. Kosmos linked every sentence in a report to the code that ran or to a paper source. Without such a trail, checking what an agent produced would mean people redoing all the research from scratch.
• Measure accuracy separately for each kind of claim. Edison’s evaluation showed that the agent is good at data analysis but weak at interpretation. If you look at only one overall accuracy score, you may miss this weakness.
• Design clearly at which stage people step in. The early version of Kosmos was built to run for hours alone, with no human involvement, and as a result it wasted heavy computation even on simple questions. Edison redesigned the system so that people can step in with feedback during the research. How often people can step in is not a side feature. It is a design choice.

What has not yet been fully proven

A large part of the data and results in this article rests on announcements by Edison and its partners. Kosmos’s accuracy figure comes from evaluating 102 statements pulled from three reports. The figures that it did six months’ or 4.1 months’ worth of research are also estimates by collaborators and by the company.

The comments from outside scientists in Alzforum and preLights were all written after trying the early version, before the overhaul. For the version rebuilt in May 2026 as a continuous collaborator, no independent outside user reports could be found yet, and the results at Population Health Partners also rest on a case study Edison wrote.

The original Nature paper on the earlier system Robin could not be checked directly. Robin’s validation experiments were done at the preclinical stage with living cells, and the researchers themselves note that the result comes from a single drug-discovery project.

Sources

1. Edison Advances — Kosmos: An AI Scientist for Autonomous Discovery — https://advances.edisonscientific.com/research/announcing-kosmos/ — 2025-11-05
2. arXiv — Kosmos: An AI Scientist for Autonomous Discovery (Mitchener et al.) — https://arxiv.org/abs/2511.02824 — 2025-11-04
3. Edison Advances — How We Built Kosmos — https://advances.edisonscientific.com/research/how-we-built-kosmos — 2026-05-18
4. Edison Scientific — Building AI-Native Biopharma: Announcing Our Partnership with Incyte Corporation and Kosmos for R&D Teams — https://edisonscientific.com/news/building-ai-native-biopharma-announcing-our-partnership-with-incyte-corporation-and-kosmos-for-r-d-teams — 2026-05-19
5. Edison Scientific — How Population Health Partners accelerates its regulatory process with Kosmos — https://edisonscientific.com/news/how-population-health-partners-accelerates-its-regulatory-process-with-kosmos — 2026-09-30
6. Edison Scientific docs — Overview — https://docs.edisonscientific.com/ — accessed 2026-10-22
7. Edison Scientific docs — Best practices for optimizing Kosmos workflows — https://docs.edisonscientific.com/guides/best-practices-for-optimizing-kosmos-workflows.md — accessed 2026-10-22
8. FutureHouse — Demonstrating end-to-end scientific discovery with Robin: a multi-agent system — https://futurehouse.org/research-announcements/demonstrating-end-to-end-scientific-discovery-with-robin-a-multi-agent-system — 2025-05-20
9. arXiv — Robin: A multi-agent system for automating scientific discovery (Ghareeb et al.) — https://arxiv.org/abs/2505.13400 — 2025-05-19
10. Edison Advances — Robin: A multi-agent system for automating scientific discovery — https://advances.edisonscientific.com/research/robin-multi-agent-system — 2026-05-19
11. SiliconANGLE — Edison Scientific raises $70M to build autonomous AI scientists for research — https://siliconangle.com/2025/12/18/edison-scientific-raises-70m-build-autonomous-ai-scientists-research — 2025-12-18
12. Alzforum — Introducing Kosmos: ‘AI Scientist’ That Makes Discoveries Overnight — https://www.alzforum.org/news/research-news/introducing-kosmos-ai-scientist-makes-discoveries-overnight — 2025-11-21
13. preLights — Kosmos: An AI Scientist for Autonomous Discovery (Manuel Lessi, Roberto Amadio) — https://prelights.biologists.com/highlights/kosmos-an-ai-scientist-for-autonomous-discovery/ — 2026-02-04