--- title: Morgan Stanley's ALPHALAB: Multi-Agent Research Across... description: Morgan Stanley's AI research team presents ALPHALAB, a multi-agent agentic harness designed to automate quantitative research. The system operates in three... canonical: https://daily.dev/posts/morgan-stanley-s-alphalab-multi-agent-research-across-optimization-domains-brendan-rappazzo-cj3mt3jrw twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Morgan Stanley's ALPHALAB: Multi-Agent Research Across Optimization Domains — Brendan Rappazzo | daily.dev og:description: Morgan Stanley's AI research team presents ALPHALAB, a multi-agent agentic harness designed to automate quantitative research. The system operates in three... og:url: https://daily.dev/posts/morgan-stanley-s-alphalab-multi-agent-research-across-optimization-domains-brendan-rappazzo-cj3mt3jrw og:image: https://api.daily.dev/og/posts/cJ3mT3Jrw.png og:image:alt: Morgan Stanley's ALPHALAB: Multi-Agent Research Across Optimization Domains — Brendan Rappazzo og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Morgan Stanley's ALPHALAB: Multi-Agent Research Across Optimization Domains — Brendan Rappazzo **[AI Engineer](https://daily.dev/sources/aidotengineer)** · 20 min read · 2 upvotes · 0 comments ## Summary Morgan Stanley's AI research team presents ALPHALAB, a multi-agent agentic harness designed to automate quantitative research. The system operates in three phases: a research phase where agents build context and form hypotheses, an eval-building phase using critic agents to validate evaluation code, and a mass experimentation phase using a strategist-worker agent architecture modeled on a Kanban board. Built without off-the-shelf agent frameworks, it uses shell access, web search, and GPU cluster orchestration (Slurm) as core tools. Initial results include top 12% placement in an Nvidia Kaggle competition and meaningful internal model improvements heading to production. The team identified key failure modes and is now building ALPHALAB 2.0, which focuses on rigorous environment and eval design as the primary mechanism for encoding domain expertise and enabling self-recursive harness improvement via meta-optimization and RL fine-tuning of open-source models. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### What are the three phases of Morgan Stanley's AlphaLab agentic research harness? AlphaLab runs research, evaluation-building, and mass-experimentation phases. The research phase builds context and a to-do list over roughly 3-4 hours using web search and shell access; the eval-building phase uses a builder agent plus two critic agents to check for conceptual errors and write unit/integration tests; the experimentation phase uses a strategist agent and worker agents organized as a Kanban board to submit and analyze Slurm-managed training jobs. _Teams designing multi-agent research pipelines can follow ongoing coverage of agent harness architectures on daily.dev._ ### How did Morgan Stanley's AlphaLab agent perform on the Nvidia NeMoTron reasoning model Kaggle competition? AlphaLab placed in the top 12% of submissions in a Kaggle competition hosted by Nvidia to fine-tune the NeMoTron model into a reasoning model. It only had about 10 iterations to work with because the team joined late, and the harness is believed to perform better with more iterations to explore. _Developers evaluating AI research agents against competitive benchmarks can track results like this on daily.dev._ ### Why did Morgan Stanley decide to build its own agent harness instead of using an off-the-shelf agent framework for AlphaLab? Building the harness from scratch, largely written by Claude itself, gave maximum freedom to tweak any part of the system and kept it provider-agnostic through functional tool calling, so it could switch easily between OpenAI, Anthropic, or open-source models. This avoided being locked into a framework's assumptions as models rapidly improved. _Engineers weighing custom agent harnesses against existing frameworks can follow similar build-vs-buy debates on daily.dev._ ## Similar posts on daily.dev - [Google I/O showed how the path for AI-driven science is shifting](https://daily.dev/posts/google-i-o-showed-how-the-path-for-ai-driven-science-is-shifting-p08sxf2im) · MIT Technology Review · 0 upvotes · 0 comments - [The Multi-Agent Trap](https://daily.dev/posts/the-multi-agent-trap-gcmbn0u9m) · Towards Data Science · 1 upvotes · 0 comments - [Anthropic’s Designs Three-Agent Harness Supports Long-Running Full-Stack AI Development](https://daily.dev/posts/anthropic-s-designs-three-agent-harness-supports-long-running-full-stack-ai-development-oqw4dgvdn) · InfoQ · 2 upvotes · 0 comments --- Tags: [#machine-learning](https://daily.dev/tags/machine-learning), [#ai-agents](https://daily.dev/tags/ai-agents), [#reinforcement-learning](https://daily.dev/tags/reinforcement-learning) [View this post on daily.dev](https://daily.dev/posts/morgan-stanley-s-alphalab-multi-agent-research-across-optimization-domains-brendan-rappazzo-cj3mt3jrw) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Morgan Stanley's ALPHALAB: Multi-Agent Research Across Optimization Domains — Brendan Rappazzo","url":"https://daily.dev/posts/morgan-stanley-s-alphalab-multi-agent-research-across-optimization-domains-brendan-rappazzo-cj3mt3jrw","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/morgan-stanley-s-alphalab-multi-agent-research-across-optimization-domains-brendan-rappazzo-cj3mt3jrw"},"datePublished":"2026-07-29T17:21:51.063Z","dateModified":"2026-09-14T06:51:49.220Z","description":"Morgan Stanley's AI research team presents ALPHALAB, a multi-agent agentic harness designed to automate quantitative research. The system operates in three...","image":"https://i.ytimg.com/vi/kiqubc5b5Yo/sddefault.jpg","thumbnailUrl":"https://i.ytimg.com/vi/kiqubc5b5Yo/sddefault.jpg","isAccessibleForFree":true,"articleSection":"AI Engineer","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"AI Engineer","logo":"https://media.daily.dev/image/upload/s--u5PucxNT--/f_auto/v1724338940/logos/aidotengineer","url":"https://daily.dev/sources/aidotengineer"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/morgan-stanley-s-alphalab-multi-agent-research-across-optimization-domains-brendan-rappazzo-cj3mt3jrw","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":2},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"machine-learning,ai-agents,reinforcement-learning","timeRequired":"PT20M","video":{"@type":"VideoObject","name":"Morgan Stanley's ALPHALAB: Multi-Agent Research Across Optimization Domains — Brendan Rappazzo","description":"Morgan Stanley's AI research team presents ALPHALAB, a multi-agent agentic harness designed to automate quantitative research. The system operates in three...","thumbnailUrl":"https://i.ytimg.com/vi/kiqubc5b5Yo/sddefault.jpg","uploadDate":"2026-07-29T17:21:51.063Z","duration":"PT20M","url":"https://api.daily.dev/r/cJ3mT3Jrw","embedUrl":"https://www.youtube.com/embed/kiqubc5b5Yo"}} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"AI Engineer","item":"https://daily.dev/sources/aidotengineer"},{"@type":"ListItem","position":3,"name":"Morgan Stanley's ALPHALAB: Multi-Agent Research Across Optimization Domains — Brendan Rappazzo"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/morgan-stanley-s-alphalab-multi-agent-research-across-optimization-domains-brendan-rappazzo-cj3mt3jrw#faq","mainEntity":[{"@type":"Question","name":"What are the three phases of Morgan Stanley's AlphaLab agentic research harness?","acceptedAnswer":{"@type":"Answer","text":"AlphaLab runs research, evaluation-building, and mass-experimentation phases. The research phase builds context and a to-do list over roughly 3-4 hours using web search and shell access; the eval-building phase uses a builder agent plus two critic agents to check for conceptual errors and write unit/integration tests; the experimentation phase uses a strategist agent and worker agents organized as a Kanban board to submit and analyze Slurm-managed training jobs. Teams designing multi-agent research pipelines can follow ongoing coverage of agent harness architectures on daily.dev."}},{"@type":"Question","name":"How did Morgan Stanley's AlphaLab agent perform on the Nvidia NeMoTron reasoning model Kaggle competition?","acceptedAnswer":{"@type":"Answer","text":"AlphaLab placed in the top 12% of submissions in a Kaggle competition hosted by Nvidia to fine-tune the NeMoTron model into a reasoning model. It only had about 10 iterations to work with because the team joined late, and the harness is believed to perform better with more iterations to explore. Developers evaluating AI research agents against competitive benchmarks can track results like this on daily.dev."}},{"@type":"Question","name":"Why did Morgan Stanley decide to build its own agent harness instead of using an off-the-shelf agent framework for AlphaLab?","acceptedAnswer":{"@type":"Answer","text":"Building the harness from scratch, largely written by Claude itself, gave maximum freedom to tweak any part of the system and kept it provider-agnostic through functional tool calling, so it could switch easily between OpenAI, Anthropic, or open-source models. This avoided being locked into a framework's assumptions as models rapidly improved. Engineers weighing custom agent harnesses against existing frameworks can follow similar build-vs-buy debates on daily.dev."}}]} ```