In 2024 our team won first place at the AGI House × Sundai.club AI Agents Hackathon at MIT with Society of Scientists — a multi-agent system that turns a research prompt into a structured, interdisciplinary grant proposal. The idea sits in the same research lineage as SciAgents (Ghafarollahi & Buehler, MIT): graph- and society-style reasoning for scientific discovery, but aimed at the painful early stage where a PI has intuition and PDFs, not yet a full narrative. The living codebase is public at github.com/Caerii/SocietyofScientists (a fork extending SciAgentsDiscovery); the README there is the ground truth for install, agents, and docs.
What the system does
You give a topic; a structured group chat runs between specialized roles: domain-oriented scientists (e.g. vision, language models, AI hardware), a planner that turns their contributions into an outline, grant writers who flesh out hypothesis, methods, objectives, budget, ethics, novelty, and prior work, and a critic that reviews and pushes back. The output is meant to be edited by a human — not filed untouched — but it is a full proposal-shaped artifact, not a bullet list.
Stack (as in the repo today)
Orchestration is built on AG2 (the AutoGen lineage) with AI21 Jamba models for long-context turns. Exa backs live search when configured; without an Exa key, the system can fall back on cached research summaries (hundreds of papers across domains). There is a Python package (society_of_scientists), a FastAPI API surface, optional frontend, scripts, tests, and built-in cost tracking so runs stay interpretable in dollars and tokens.
Agents and documentation
The repository documents twelve agents and their roles, the end-to-end generation pipeline, configuration (AI21_API_KEY, optional EXA_API_KEY), and architecture notes. For citations and relationship to the original SciAgents paper, see docs/README_SciAgents.md and the attribution block in the upstream README — the hackathon demo was the wedge; the repo is where the system kept growing.
Why it mattered to us
Scientific funding is a bottleneck; the hardest part is often the phase before the budget spreadsheet — figuring out what is novel, what is crowded, and how to phrase a bet. If agentic tooling can shorten that without confidence, it lands in the same ethical bucket as copilots for code: high leverage, high responsibility. The hackathon format forced us to pick one vertical slice and prove it live.
Team and judges
Teammates Igor Sadalski, Elior Benarous, and Alif Jakir (with Dünya Baradari in the public thread celebrating the win). Judges Tuan Ho, Mark Weber, and Yaniv Markovski — thank you for the time and feedback.
What I contributed
I worked across integration, prompting, and demo narrative so the multi-agent graph stayed coherent under time pressure — the same “glue engineer” role I tend to take when half the team is modeling and half is pitching. The GitHub history on SocietyofScientists reflects continued hardening after the event (docs, API, tests, bugfixes). I can still add a more personal retrospective here later.
Attribution
The project extends ideas in SciAgents: Automating Scientific Discovery Through Bioinspired Multi-Agent Intelligent Graph Reasoning (Ghafarollahi & Buehler, Advanced Materials, 2024). The repo README and docs/README_SciAgents.md spell this out precisely; I cite it here so the intellectual debt is visible on my site, not only on GitHub.
Links
- github.com/Caerii/SocietyofScientists — source, docs, quick start (
pip install -e .) - LinkedIn — team shout-out and hackathon summary
MIT Reality Hack 2025 — OpenGalea · MIT Reality Hack 2026 — Personoid