This repository is the public catalog and governance backbone for Microsoft Discovery, a platform of containerized scientific AI agents for research domains such as chemistry, materials, bioinformatics, and EDA. The dominant code is not a single runtime but a monorepo of agent definitions (metadata.yaml, agent.yaml, tool.yaml), tool entrypoints, Dockerfiles, and a large GitHub Actions validation framework that enforces structural, schema, policy, tagging, and security rules on PRs. It also includes an Azure 'supercomputer' CLI for deploying tool definitions, storage assets, and polling job runs, plus agent evaluation utilities. The catalog is the central asset; actual agent containers are built from Dockerfiles in this repo and run with a standard /input -> /output + final_results.json contract. A secondary but real data-research strand exists: the gwp-predictor agent includes a full training pipeline (IPCC/Hodnebrog data ingestion, SMILES resolution, Tanimoto split, Chemprop ensemble training, active learning, benchmarking), and several other agents embed scientifically meaningful utilities (RDKit hazard-group SMARTS, BindingDB SQLite builder, PDB parsing). These are packaged as agent tools rather than a standalone data platform, so the repo is classified as an application with hybrid research content. The strongest engineering is in the governance layer: a policy rule engine with Rule/Finding/Severity/Scope abstractions, waivers, ratchet baselines, and fail-closed config; a dependency-free Dockerfile parser; content sniffers for binaries and model weights; and a supply-chain hardened package verifier embedded in one agent. These components are isolated enough to be extracted. The main weaknesses are heavy duplication of generated FlowBlind runtime files across three tool directories, hardcoded container paths, unpinned base images, and some tests tightly coupled to Azure CLI stderr text.