AI Agent

An AI agent takes a goal and works toward it over many steps. It decides which tool to call, reads the output, adjusts its plan, and continues. Examples: Claude Code resolves GitHub issues over dozens of edits, Cursor's Composer refactors across files, Operator browses the web and fills forms, Manus orchestrates research tasks. Agents need three things: planning, tool use, and memory across steps.

Agent architectures vary: ReAct (reason + act interleaved), function-calling loops, tree-search planners, and hybrid designs. The dominant 2026 pattern is a reasoning model (o3, Claude Opus with Extended Thinking, DeepSeek R1) plus tool definitions via function-calling or MCP. The agent runs in a loop: reason about next step → call tool → observe result → reason again → ... until done. State management includes conversation history, file system state, and task decomposition. Safety-critical: agents can make irreversible changes (send emails, modify files, spend money). Most production agents require explicit user approval for high-risk actions.

Formal frame: an agent is a policy π(a | s, g) mapping state + goal to actions. In LLM agents the policy is the LLM + a harness that parses tool calls, executes them, and feeds results back. Planning depth is bounded by effective context length; long-horizon agents use hierarchical task decomposition to stay within context. Error recovery: agents that can detect failed tool calls and replan outperform linear-execution agents by 30-50% on SWE-bench Verified. Reflexion, Self-Refine, and Tree-of-Thought extend single-shot reasoning into multi-iteration improvement. Cost is the dominant constraint · agent runs can consume 100K-10M tokens per task.