Revisiting the End-to-End Design Principle for An Internet-of-Agents
The excitement about a futuristic Internet-of-Agents is nevertheless tempered by reality checks on real-world reliability of agents. Existing efforts in the ML community around agentic deployments largely focus on taxonomies of failures, offline benchmarks, and LLM-based evaluations in closed settings. Drawing upon decades of wisdom from the Internet evolution and the end-to-end design principle that has stood the test of time, we posit that such approaches are fundamentally insufficient if the agentic hype needs to live up to its potential. Analogous to successful deployments of “over the top” applications such as Internet video over an unreliable network substrate, we envision agentic systems will need a management plane driven by in situ, full census, measurements of perceptual reliability in the field rather than offline benchmarks, evaluations, and tests. We present an early proof-of-concept and discuss open challenges.
Read More
