There will soon be AI agents working on our behalf (blog.cip.org)

AI’s next phase will be multi-agentic: instead of one monolithic model, capability will emerge from networks of specialized agents coordinating in parallel. Architectures like Mixture-of-Experts and multi-agent platforms (AutoGen, CrewAI) already show gains in efficiency and problem-solving that single models can’t match; Anthropic, OpenAI (IMO solution), and Sakana provide concrete examples. This shift changes the bottleneck from raw model intelligence to coordination: existing schedulers, consensus protocols, and human-centric governance were built for deterministic, human-limited systems and won’t scale to millions of non‑deterministic agents. A critical technical danger is compounding error—tiny misalignments can cascade through exponential agent interactions—so post-hoc feedback won’t be enough to keep systems aligned. A proposed solution is a “representative agent”: a persistent, adaptive model of a person’s preferences, goals, and values that serves as a continuous alignment signal and arbiter (e.g., personalized forecasts for policy impacts). Building such agents raises hard problems—eliciting high-dimensional values, deciding when to act autonomously versus defer, and resisting adversarial manipulation. Trust requires rigorous evaluation: frameworks must measure fidelity, pass a “Volitional Turing Test,” and dynamically probe for gaps. The Collective Intelligence Project is pursuing this via Global Dialogues (large-scale ground-truth surveys) and Weval (context-specific evaluations) to benchmark LLM fidelity and inform protocols, tools, and collective-intelligence systems for a future of delegated AI.