Researcher proposes agent-suggests-human-executes loop for real-world science experiments
suragnair · x · 2026-09-25
Researcher suragnair outlined a staged scheme for letting AI agents participate in real scientific experiments:
- Collaboration loop: give the agent a science problem, it proposes the next step, a human executes it (rather than a machine), and the cycle repeats until a verifiable final output — e.g., a biology experiment can be checked via QC results.
- Full-observability error attribution: with sensors and cameras at scale, failures can be attributed to agent error, human error, or environment error (e.g., high temperature), letting agents learn how everything interacts and improve experiment execution over time.
In discussion with Stanford professor Anshul Kundaje, both agreed that since physical-world missteps could damage equipment or harm operators, the executors should initially be human experts who intervene on hazardous actions — effectively a safety oracle for the agent.
More from coding & agent
- Stream launches Agent Skills so coding agents can build chat, video and moderation apps from one prompt — amos_gyamfi · 2026-09-25
- NVIDIA Open-Sources Nemotron-Terminal: Boosts Qwen3-32B from 3.4% to 27.4% on Terminal-Bench 2.0 — _weiping · 2026-09-25
- 5 Ways to Handle Conflicting Information in AI Systems, From Source Ranking to Surfacing Disagreement — goyalshaliniuk · 2026-09-25
- TypeLLM adds type-safe, schema-guaranteed generation to LLMs without touching weights — kalyan_kpl · 2026-09-25
- kuberdenis talks harness engineering and dev workflows with alxfazio, visuals made by Opus 5.5 — tensorqt · 2026-09-25
- Five mistakes people make with agentic factories: micromanaging, no quality bar, and 'puppet shows' — hugobowne · 2026-09-25