MSDINTERNATIONAL← Home Discover MSD ⌄From AI and robotics to islands, yachts, specialist partners and new business models. Choose your entry point – and discover what lies behind it.
AI, robotics & futureFuture · Die Firma aus Systemen10,000 AI agents as a scaling target, networked tasks and an interactive estimator.
↗
AI, robotics & futureRobotics · 10 platformsManufacturers, capabilities, interfaces and integration tasks.
↗
MSD Direct & MarktplatzMSD Direct · Das PortalPeople, opportunities, way of life.
↗
MSD WORLDSiteWarden · Robotics LabGripping, sensor technology and physical AI – planned experimental programs.
↗Die Übersicht wird geladen.
MSD / Research / RESEARCH EDITION
Reflection examines agents who use feedback in linguistic form for later experiments. The feedback changes the work context.
Jetzt eintauchen ↓RECHERCHESTAND 12.11.2026 · FAKTEN + EINORDNUNG + SZENARIENKI & Agenten
External research · MSD classification
Keine Behauptung eines eigenen Versuchs.
01 / EINORDNUNG
Plausible self-criticism is not independent evidence of quality; incorrect feedback can reinforce errors.
In the case of learning loops, keep an external test and unchanged comparison tasks.
This derivation is an editorial recommendation from MSD. It is not a further result of the study cited.
Shinn et al. · Reflexion · Publikations-/Bezugsjahr 2023. Recherchiert am 12.11.2026.
Read primary source ↗External research, not an MSD customer reference. We will only publish our own experiments with their setup, date, measurements and documented failed attempts.
02 / LINKS
In a randomised trial with 16 experienced developers and 246 tasks, early-2025 AI tools increased completion time by 19%. Participants perceived their benefit much more favourably.
METR · Randomised trialMETR describes a follow-up experiment in February 2026. Evidence of faster work is difficult to interpret due to selection of participants and tasks as well as problems with time measurement.
METR · Methodenupdate 24.02.2026The Transformer paper presents an attention-based architecture for sequence processing. It is a methodological building block of later language and multimodal models.
Vaswani et al. · Attention Is All You Need