The Challenge of Managing Agentic AI
Tatyana Mamut of The Roll-Up Co. argues that managing agentic AI is analogous to leading highly intelligent employees. She notes, "You want to give AI agents as much autonomy as possible. But this is just like really smart employees... The autonomy-control balance is the defining challenge of agentic AI." This framing highlights the emerging leadership skills required to manage increasingly autonomous systems.
The core challenge is shifting from task-specific robots to generalist agents powered by foundation models. Companies like Covariant, founded by OpenAI and UC Berkeley alumni, are building these models for physical world applications, aiming for the same transformative impact seen with LLMs like GPT. NVIDIA's Project GR00T is a major initiative to create an open foundation model specifically for humanoid robots, enabling them to understand multimodal instructions and learn from human demonstrations. This technical shift directly impacts defense strategy, where the Pentagon is pursuing "mass" through autonomous systems to counter adversaries. The DOD's Replicator Initiative aims to field thousands of "attritable" autonomous systems across multiple domains by August 2025. Replicator-2, announced in September 2024, specifically targets the growing threat of small counter-UAS systems, demonstrating a focus on both deploying and defeating agentic swarms. For engineering leaders, the focus is on creating software-first platforms that can command diverse hardware. Anduril's Lattice software is a key example, serving as an AI-driven command and control system that integrates disparate sensors and autonomous platforms. The U.S. Army recently selected Lattice as the fire control platform for its counter-UAS missions, highlighting the demand for open-architecture systems that can rapidly integrate new capabilities. This approach allows a single human operator to manage teams of robotic assets, a critical force multiplier. In the commercial sector, humanoid robots are moving from R&D into pilot deployments. Figure AI signed a commercial agreement with BMW to deploy its humanoids in the Spartanburg, South Carolina manufacturing plant for tasks deemed difficult, unsafe, or tedious. While this is a milestone, the initial phase focuses on identifying use cases rather than a full-scale rollout, underscoring the careful, milestone-based approach to integrating these systems. Agility Robotics has also deployed its Digit humanoid in logistics operations with partners like GXO and Amazon. Venture funding in robotics has become highly concentrated, with fewer startups raising larger rounds. In 2024, robotics startups raised $6.4 billion, with significant rounds for Figure ($675M) and Physical Intelligence ($400M). Investors like Jeff Bezos's Bezos Expeditions are backing companies focused on AI-powered, versatile robots. This trend signals a market shift away from single-purpose machines toward generalist, adaptable hardware. Foundational research in robot learning is accelerating the push toward generalist capabilities. Google DeepMind's ALOHA 2, an open-source bimanual teleoperation system, is enabling the collection of large-scale demonstration data needed to train dexterous policies. This research is proving that imitation learning, when combined with large datasets and expressive models, can teach robots complex, contact-rich tasks like tying shoelaces, which were previously considered extremely difficult.
