In his talk, Rajiv Chandegra, a medical doctor and AI engineer at Annicha Labs, introduces a new design philosophy for AI engineering called adaptive engineering. He begins by explaining the current paradigm, where AI agents are guided by fixed, pre-engineered harnesses—structured frameworks that define roles, tools, and sequences ahead of runtime. These harnesses, such as Claude Code or LangChain, provide reliability, auditability, and predictability, making them well-suited for complicated but static problems. However, Chandegra argues that as AI models become more powerful and begin interacting with the dynamic, messy real world, these fixed harnesses will become brittle and outdated.
Chandegra contrasts two worldviews: the reductionist or analytical view, which sees the world as composed of stable parts that can be individually analyzed, and the systems or relational view, which emphasizes relationships and processes as fundamental. He highlights that many real-world problems are complex systems characterized by diverse agents interacting locally, adapting continuously, and producing emergent behaviors that cannot be predicted by analyzing parts alone. Examples include flocks of birds or the wetness of water—phenomena arising from relationships rather than individual components. This complexity challenges the factory-like fixed harness approach, which assumes decomposable, predictable problems.
He distinguishes between complicated problems, which are knowable and predictable (like building a jumbo jet), and complex problems, which involve adaptive, interacting agents and emergent phenomena (like markets or organizations). The fixed harness approach works well for complicated problems but fails in complex, real-world scenarios where continuous adaptation and emergence are necessary. Chandegra proposes adaptive engineering as a new discipline where the harness is not fixed but emerges and evolves during runtime, shaped by agent interactions and environmental feedback, allowing the system to self-organize and adapt dynamically.
In adaptive engineering, the engineer’s role shifts from prescribing fixed structures to designing constraints or “rules of the game” that enable agents to interact, learn, and specialize autonomously. The harness becomes an emergent, decentralized, self-organizing multi-agent system that continuously adapts to changing conditions. Chandegra emphasizes that this approach is not about abandoning engineering but relocating its focus to sensing and responding to emergent structures rather than controlling them rigidly. He contrasts this with existing systems like Hermes AI, which focus on vertical intelligence (improving individual agents), whereas adaptive engineering prioritizes horizontal intelligence—coordination among multiple agents.
Finally, Chandegra acknowledges the challenges and risks of adaptive engineering, including potential stability on suboptimal attractors, loss of legibility, unpredictability, and risks of monoculture among agents. Despite these, he believes adaptive engineering is essential as AI moves into real-world, multi-agent, multi-institutional environments where fixed harnesses cannot keep pace. He calls for further exploration of this paradigm, emphasizing that the future of AI engineering lies in harness adaptability and decentralized coordination, enabling AI systems to operate effectively in complex, evolving real-world contexts.
