Agentic AI Systems (2024) › Applications and Impact

AI coding assistants illustrated how agents could plan and implement complex software features by breaking requirements into smaller tasks and using development tools. An agent could inspect a codebase, identify the required change, implement it, run tests, and iterate on feedback with limited human intervention.

AI research assistants could autonomously conduct literature reviews, search for relevant papers, analyze findings, synthesize information, plan experiments, and even write and submit research papers. These agents could manage the entire research workflow from initial exploration through publication, coordinating across multiple tools and systems to accomplish complex research tasks.

AI personal assistants could manage complex workflows, coordinating across multiple applications and services to complete tasks that previously required human intervention. Agents could schedule meetings by checking calendars, sending emails, coordinating with participants, and handling follow-ups. They could manage travel by researching options, comparing prices, making reservations, and handling logistics.

Agentic behavior expanded what AI assistants and automation systems could do. Systems that initiated work and executed plans could support new forms of human-AI collaboration. Applications appeared in healthcare and education, along with business process automation.

In healthcare, agentic systems could assist with patient care coordination across several providers and information systems. An agent could coordinate appointments and care plans while sending medication reminders or patient messages.

In education, agentic AI could support personalized learning by adapting instruction based on student progress, managing learning resources, coordinating assignments, and providing individualized feedback. These agents could maintain awareness of each student's learning state and adapt teaching strategies accordingly.

In business process automation, agentic systems could manage workflows that combined decisions with coordination and adaptation. Unlike a fixed rule-based process, an agent could handle work requiring context-dependent judgment and flexibility, tasks previously assigned to people.

The architectural principles established by agentic AI systems also influenced other areas of machine learning and AI. The ideas of persistent memory, tool use, and autonomous planning were applied to other types of systems, including robotics, autonomous vehicles, and smart home systems. The techniques developed for safe execution and alignment were also adapted for other applications that required AI systems to interact with the real world.

The implications of agentic AI extended far beyond individual applications to broader questions about the future of work and human-AI collaboration. Agentic systems could potentially automate many tasks that had previously required human intelligence and creativity, raising questions about the future of employment and the nature of human work. At the same time, they could also augment human capabilities, allowing people to focus on higher-level tasks while AI systems handle routine and repetitive work.

Agentic systems also required new evaluation methods. Single-turn benchmarks could not measure whether a system developed and carried out a plan, then adapted over an extended period. New frameworks therefore assessed long-term task performance and safety.