Summary
Brendan Rappazzo from Morgan Stanley’s AI research group presents AlphaLab, an internal multi-agent system built to automate quantitative research — from initial hypothesis formation through evaluation construction and mass experimentation — at the AI Engineer conference.
The talk opens with context on the team: roughly 30 PhD AI researchers operating half as an academic group (publishing, open-sourcing) and half on applied internal work. The framing problem is that many quant research tasks have a Kaggle-like structure — given a time series input, optimize predictions subject to calibration constraints — making them well-suited for automation once models can handle long-horizon coding tasks. Rappazzo marks December 2025 and the release of Claude Opus 4.5 in agentic harnesses like Claude Code as the inflection point that made this feel tractable.
AlphaLab’s architecture runs in three phases: a research phase where the agent builds context by reading documentation, running statistical tests, and querying arXiv; an eval-building phase using a multi-agent critic loop (one builder, one high-level conceptual critic, one programmatic unit-test critic) that iterates until all agents agree the evaluation is sound; and a mass experimentation phase formulated as a Kanban board where a strategist agent continuously generates and queues new experiments. The v2.0 changes highlighted include improved eval construction robustness, GPU cluster job submission via natural language config, and dynamic context management across agents. The full AlphaLab 1.0 technical report and code are publicly available on GitHub.
📺 Source: AI Engineer · Published July 29, 2026
🏷️ Format: Keynote Launch
