Constructor Talks: The AI Scientist in the Wild: What a Hundred-Plus Autonomous Research Agents Teach Us About the Future of Discovery
“AI Scientist” systems — agents that generate hypotheses, run experiments, and write papers — have moved from thought experiment to a crowded ecosystem in under three years. But beneath the demos and headlines, what has this field actually learned? This talk presents the results of a large-scale meta-review of the autonomous-research-agent landscape: over a hundred open-source systems and their accompanying papers, systematically mapped from GitHub and arXiv into distinct “schools” of thought, recurring design patterns, and — just as importantly — recurring anti-patterns.
From this corpus, a small set of empirical regularities emerges: how these systems actually improve, where their evaluation practices break down, which capabilities remain systematically missing, and why a large “silent stratum” of projects never reports failures at all. The talk distills these regularities into a practical map of where autonomous research is genuinely working, where it is stuck, and what that means for anyone planning to build on it.
The closing part turns from observation to design: if agents are to do real science, they need explicit structure to coordinate around. I will show two directions we pursue at Constructor Knowledge Labs: Semantic Flow, which makes the reasoning behind generated scientific and educational content machine-actionable and auditable, and an “article-as-specification” workflow, where a future paper is unfolded backwards into claims, evidence, and tasks — so that human researchers and AI agents can work as one team toward results neither could reach through the usual writing process.
- Date & Time: 2 October, 13:00–14:00 CEST
- Format: On-site and online
- Location: IRC Conference Room (Constructor University, Bremen)
- Link to join online: https://constructor-org.zoom.us/j/95097566642?pwd=cFmlBR3vTbqvp6kPDhSnDaPfiShp9R.1 (Meeting ID: 950 9756 6642, Passcode: 774314)
No registration is required, and participation is free of charge. External guests are very welcome to attend the Constructor Talks, both on-site and online.
Speaker: Prof. Dr. Andrey Ustyuzhanin
Principal Investigator at Constructor Knowledge Labs, Adjunct Professor of Computer Science at Constructor University Bremen, Director of AI Research at Constructor Knowledge, working at the intersection of agentic AI systems, software engineering, and scientific workflows.
Moderator: Dr. Mariia Snigireva, Managing Director of Constructor Labs
Agenda:
- 12:45–13:00 – Doors and stream open
- 13:00–13:05 – Welcome and Introduction
- 13:05–13:50 – Keynote Presentation by Prof. Dr. Andrey Ustyuzhanin
- 13:50–14:00 – Discussion and Q&A Session followed by informal discussion over coffee
Topics Covered in the Session
- The AI Scientist landscape in 2026: a systematic map of 100+ autonomous research agents from GitHub and arXiv, organised into schools of thought.
- Six empirical regularities: how self-improving research agents actually behave, and the anti-patterns they keep repeating.
- Why evaluation is the weakest link: benchmark escape, missing cost accounting, and the “silent stratum” of unreported failures.
- From plausible text to auditable reasoning: making agent output verifiable with explicit reasoning structures (the Semantic Flow approach).
- The research paper as an executable specification: coordinating humans and AI agents around claims, evidence, and tasks — a glimpse of how research teams may work next.
Who Should Attend
- Researchers, PhD students, and research team leads who want a rigorous, evidence-based picture of autonomous research agents — and a preview of how human–AI research teams may be organized.
- R&D and innovation leaders deciding whether and where to deploy AI-assisted research pipelines in their organisations.
- Educators and knowledge-product builders interested in auditable, reasoning-driven content generation.
Constructor Labs aims to keep the session open and discussion-driven, with space for different perspectives and honest exchange.