Autonomous Agents for Scientific Tasks
summary
Claim: autonomous agents improve on real scientific measurement tasks when they generate and test explicit hypotheses. Demonstration: an insilicopet CT→PET pipeline built at Radicait where moving from a 2.5D channel-stacked generator to 3D convolutions produces a clear step change in performance. Method: scaffolded hierarchy of documents in Obsidian plus adversarial/collaborative agent loops generate hundreds of alternative designs, use Oracle CLI with GP5.5 Pro and multimodal reviewers, and show measurable gains on industrial datasets.
AI-written summary · about
chapters (9)
- 0:00Autonomous agents for science
- 1:04Research plateaus with current agents
- 3:52Decomposing scientific problems
- 6:41Generating hypotheses hierarchically
- 9:49Improving models with loops
- 12:18Adversarial collaboration loops
- 15:46Multimodal skills for agents
- 17:13Future challenges for agents
- 18:59Summary
related talks
41mGoing beyond copilot with AI Agents
28mAgentic AI
34mAI Agents: Transforming Autonomous Decision-Making
30mBuilding Reliable AI Agents: An Open Source Approach To Evaluation and Observability
26mBuilding Agentic Workflows: Patterns for Orchestrating Intelligent Systems.
29mPairing with the machine: How AI Agents are rewriting the rules of software development
49mAI Agents: Collaboration in Real-Life Applications
15mWhat’s Next in Medical AI?