Agents’ Last Exam (ALE) evaluates AI agents on professional workflows with verifiable outcomes. Led by Berkeley RDI, the project brings domain experts into benchmark creation so that evaluation reflects the tools, inputs and deliverables used in real work.
I contribute to ALE as an author and an Architecture, Engineering and Construction (AEC) task contributor. My work brings architectural modeling, computational design, building-performance analysis and visualization into the evaluation of general-purpose AI agents.

→ Evaluating Professional Work
ALE places an agent in a reproducible desktop environment with the required software, instructions and input files. The agent works through the task using the interface and terminal, and the resulting artifacts are evaluated against reference outputs. This connects agent performance to the completion of a whole workflow, including the final files it leaves behind.
→ My AEC Contribution
I contributed workflow tasks and supporting materials drawn from architectural and building-science practice. The task packages connect a design brief and starting assets to concrete deliverables, such as editable models, parametric definitions, simulation results and rendered images.
2D → 3D Reconstruction
Reconstructing building geometry from floor plans, sections and elevations. This work tests the translation of several drawing views into a coherent three-dimensional model and usable geometry exports.
Physics-Based Form-Finding
Developing structural form through simulation in Rhino and Grasshopper, including Kangaroo workflows. The task material connects a design brief and starter geometry with a parametric definition and the resulting form.
Environmental Analysis & Machine Learning
Comparing design variants with Ladybug and Honeybee, and preparing a Python surrogate-modeling workflow around EnergyPlus simulation data. These tasks combine computational setup, analysis and the interpretation of building-performance results.
Architectural Rendering
Producing views from supplied models, materials and camera information. The contribution extends evaluation from geometry and numerical analysis to visual deliverables.
→ Why This Matters
AEC workflows combine spatial reasoning with software operation and domain knowledge. An agent must interpret a brief, manage files, work across representations and produce an artifact that can be inspected. Contributing these workflows connects my research in AI-assisted design with the practical question of how to evaluate useful design work.
→ Explore the Project
Visit the official project website for the current benchmark and results, or read the framework documentation for details of the evaluation environment. The paper and full author list describe the wider collaboration.