Heute nutzen Agenten Software. Morgen nutzen sie alles.
Wir arbeiten auf eine Welt hin, in der KI nicht nur Tasten auf Bildschirmen drückt — sondern Maschinen bedient, physische Räume navigiert und in der realen Welt mit Menschen zusammenarbeitet.
Computer-Use
Agents that see screens, click buttons, and operate any software exactly like a human would. Nein APIs needed. This is where Deck is today — and it's already changing how companies work.
LLMs innerhalb heutiger Grenzen maximieren
Current models are powerful but constrained — context windows, hallucination, cost. We're researching how to push further: multi-agent orchestration, long-running task memory, self-healing workflows that recover from errors without human intervention, and structured reasoning that keeps agents on track across 100+ step workflows.
Roboter-Use
The same principles that let an agent navigate a website can let it navigate a warehouse. The same vision model that reads a dashboard can read a factory floor. We're exploring how Computer Use extends beyond screens — into cameras, sensors, robotic arms, and physical environments.
Offene Fragen, an denen wir arbeiten
APIs cover 1% of the world's software. The other 99% has only a user interface. How do you build an agent that can operate all of it — without a single line of integration code?
If the cognitive loop for navigating a website is the same as navigating a warehouse — perceive, decide, act — what infrastructure bridges the gap between digital and physical environments?
In a 200-step workflow where each step is 98% reliable, the end-to-end success rate drops to ~2%. How do you design error correction that keeps reliability above 95% without quadratic cost?
Every enterprise has thousands of credentials and every agent needs access. How do you build a credential architecture where no single VM compromise exposes the entire graph?
Software changes its UI faster than any model can retrain. How do you build agents that treat every interface as novel — and still operate it correctly on the first attempt?
Was ist die große Frage, die Sie zu beantworten versuchen?
Wenn diese Fragen Sie begeistern