Agentic Compute
Agentic Compute is designed to turn a scientist's query into ranked hypotheses: an agent loads qualified procedures, checks rights and requests screening. Outputs stay hypotheses.
This page describes a proposed architecture. It is not evidence of an existing deployment.
How it fits the platform
It is designed to receive requests from the Interface and pass reviewable computational evidence to Scientific Review; it never approves advancement.
From query to hypotheses
A scientist submits a query, an objective and constraints. The Hermes Agent is designed to load a qualified procedure from the K-Dense-AI scientific skills, check identity, purpose and permitted use, and retrieve evidence only within the authorized scope. Permission checks are designed to run before retrieved content reaches the model.
Screening as durable workflows
The agent builds hypotheses and prepared chemical states, then submits a structured screening request to a durable workflow service. The workflow is designed to validate permissions, licenses and budget; record inputs, versions, parameters and seeds; apply bounded retries; and support cancellation and holds. Chemistry preparation and docking are designed to use RDKit, Meeko and AutoDock Vina.
Limits that stay in place
Rankings remain hypotheses, with their limitations stated alongside. Agents cannot approve advancement: scientists approve or revise the experimental work package.
Sources
- NEWMA Technology Architecture and Workflows § 2
- NEWMA technological architecture and technology stack § 2B
- NEWMA technological architecture and technology stack § 2C
See it in the demo
Opens Demo sign-in. After signing in, choose Agentic discovery (/demo/w3-agent) from the demo dashboard. In the demo this is labelled Simulated agent, Simulated workflow engine, Simulated compute.