Research paper introduces El Agente Fármaco, an AI agent for automating drug discovery workflows
The research paper describes El Agente Fármaco, an AI agent that automates computational workflows for drug discovery using a shared scientific memory called Fármaco Atlas with provenance tracking. On a 22-task benchmark, according to the authors, it outperformed plain tool calls and Codex while using fewer tokens.
Researchers published a paper on arXiv (2611.03832v1) describing El Agente Fármaco, an AI agent based on large language models that converts drug discovery goals formulated in plain language into extensive computational screening workflows. The agent itself selects scientific tools, evaluates their results, and adjusts the next steps according to the goal and the user's requirements.
A key component of the system is Fármaco Atlas, a shared scientific memory that stores molecules, structures, and results including provenance – a record of how they were obtained. Thanks to targeted queries against this database, the agent can run computations for an entire collection of molecules with a single request, without requiring the model to make a separate decision for each compound; according to the authors, this limits the amount of information the model has to process even as screening collections grow. Integrated tools support molecule generation, prediction of binding and drug-likeness properties, and searching for commercially available compounds.
According to the authors, in a retrospective comparison with an expert-led campaign, the agent proposed chemically distinct candidates with comparable docking scores and higher computed drug-likeness according to the specified selection criteria. On a benchmark of 22 tasks, it achieved the highest average score as rated by a model, compared to plain tool calling and to Codex, while having a lower median cumulative token consumption and a lower maximum context.
Why it matters
This is an academic work rather than a commercially available product, so its practical impact is mainly relevant for research teams in computational drug discovery and for developers of AI agents for scientific tasks. What's interesting is the approach using shared memory and provenance tracking, which according to the authors allows the agent to process large collections of molecules more efficiently (fewer tokens, smaller context) than general tool-calling agents such as Codex – an inspiration for agent architectures built on top of structured scientific data.
Relevant practical impact
What this means
For a business
The results show that an agent with a dedicated scientific memory and provenance record was able to handle complex multi-step tasks with lower token and context consumption than a general tool-calling approach or Codex – a relevant insight for companies developing specialized AI agents for technical or scientific workflows, not just for the pharmaceutical sector.
Development More business impacts →Check the original
Event sources
only one source so far · 1 publisher, 1 independent. We count feeds from the same owner only once.