An autonomous AI system that can perceive its environment, make decisions, use external tools (like calculators or search engines), and take actions to achieve
An AI agent wraps a model in an action loop: given a goal, it plans, calls tools (search, databases, code, email, other agents), observes results, and iterates until done or escalation. Architecture adds memory (short-term scratchpads, long-term stores), guardrails (permissions, approval gates, budgets), and tracing so every action is auditable.
Agents are 2026's dominant AI paradigm, the shift from generating answers to completing work: filing tickets, reconciling data, fixing code, coordinating processes. They concentrate both the productivity upside and the new risk surface, which is why agent design, security, and operations are the field's fastest-growing disciplines.
The LLM is the reasoning engine; the agent is the system around it: goals, tools, memory, and a control loop that acts in the world. Same model, radically different capability and risk profile.
Multiple specialized agents collaborating: a planner delegating to researcher, coder, and reviewer agents, or handoffs across vendor boundaries via protocols like A2A. It trades single-agent simplicity for specialization and parallelism.
Match autonomy to reversibility: read-only and easily-undone actions can run free; irreversible or sensitive actions (payments, deletions, external sends) get human approval gates. Autonomy is earned per action class, not granted globally.