Node Types
Hypothesis
Test
Evidence
Judgment
Update
Commitment
Select a trace from the sidebar
to visualize its epistemological graph
A framework for the science of agents and agents for science
Extensive utilities that facilitate research into agent methodologies and simplify the creation, deployment, and evaluation of scientific agents and environments.
How Corral Works
A microservice architecture ensuring flexibility, scalability, and robust isolation.
Environments
The "world" the agent interacts with. Defines the task space, available tools, and provides observable feedback. From chemistry labs to HPC clusters.
Agents
Modular entities for perception and decision-making. Built with LLMs using scaffolds like ReAct, ToolCalling, LLMPlanner, and Reflection.
Tasks
Define problems for agents to solve with scoring functions for evaluation. Chain tasks into TaskGroups for complex multi-stage challenges.
Decoupled Architecture
Corral separates agents from environments via a client-server design with REST API communication.
- CorralServer
- Hosts and manages environments, provides the interaction interface via CorralRouter.
- CorralRunner
- Executes agents, orchestrates their lifecycle, feeds observations and relays actions.
Environments
Pre-built scientific environments spanning chemistry, physics, materials science, and more.
Foundational Principles
Cite this work
If you use Corral in your research, please consider citing:
@article{ríos-garcía2026ai,
title = {AI scientists produce results without reasoning scientifically},
author = {Martiño Ríos-García and Nawaf Alampara and Chandan Gupta and Indrajeet Mandal and Sajid Mannan and Ali Asghar Aghajani and N. M. Anoop Krishnan and Kevin Maik Jablonka},
year = {2026},
journal = {arXiv preprint arXiv: 2604.18805}
}
Ready to benchmark?
Start evaluating AI agents on scientific tasks in minutes.
-
Agent Type
0
Nodes
0
Tool Calls
Load a trace directory to begin
Select a folder with JSON trace files