UNLOCK 2026 How to Build Agents For Scientific Discovery in 2026 Andrew White, Edison Scientific

Author: Andrew White, Co-founder, Edison Scientific Collection: UNLOCK-2026-Sessions Source: How to Build Agents For Scientific Discovery in 2026

Summary

Andrew White, co-founder of Edison Scientific (spun out of Future House), delivers a technically dense 18-minute talk covering the state of AI agents for scientific discovery as of 2026, drawing on three years of personal experience building these systems. He presents a timeline of milestones from his first AI agent work in January 2023 (paper QA) through Future House's founding, the development of Cosmos (a long-running scientific discovery agent), and the June 2025 spin-out of Edison Scientific. The talk establishes that model intelligence doubles every 7 months (METR benchmark data), that AI capex in 2026 is projected at $700 billion — exceeding the dot-com boom's $426 billion — and that this investment is only now beginning to show up in real-world productivity metrics like GitHub commit volume.

White argues that the architectural innovation in agents has been surprisingly stagnant since 2022 — the ReAct (Reasoning + Acting) pattern remains a strong baseline — and that the real differentiators are infrastructure quality, model training, and data rather than novel agent design. The main architectural innovation he identifies is "progressive disclosure," hiding tools in Model Context Protocol (MCP) servers that activate only when needed. For 2026, he argues the field is converging on "uni-agents" — single, general-purpose agents that write and execute code (like Claude Code) rather than many specialized agents per workflow. He introduces Project Aries, Edison's new product: a Slack-integrated persistent colleague-like agent combining the Cosmos world model with a coding agent, available in beta.

The Future House / Edison platform provides access to 175 million full-text papers, 120 million patents, and all FDA-registered clinical trials. Their Lab Bench 2 benchmark covers 1,912 questions across scientific tasks. A Cosmos-based discovery validation study recovered 7 novel scientific discoveries that were subsequently confirmed in papers, including one by Antonio's group at MIT in materials science. The platform reached 100,000 users and had over 2 million queries.