An AI coding agent is not just an LLM with a chat window. It's an autonomous system that can plan, execute, debug, and iterate on code tasks without human intervention at every step.
An LLM (like Claude Opus 4.7 or GPT-5.5) generates text. It can write code, but it cannot run it, test it, or fix errors. An agent wraps that LLM with tools — a file system, a terminal, a search engine — and decides when to use each tool to accomplish a goal.
Orchestration: The agent breaks a complex task into steps. Claude Code's Coordinator mode spawns sub-agents that work in parallel. Cursor's Composer handles multi-file edits sequentially.
Memory: Agents remember what they've done. Claude Code has 4 memory types across 5 layers. Aider uses repo-map for codebase awareness. Cline externalizes memory via MCP.
Tools: An agent's toolset determines what it can do. Claude Code has 40+ tools with code-split loading. Cursor's toolset is narrower but deeply integrated into the IDE.
Two agents using the same LLM can produce dramatically different results. The agent's architecture — how it decomposes tasks, manages context, handles errors — is often more important than the underlying model. This is why we score agents on 7 architectural dimensions at AgentRanks, not just benchmark performance.
← Back to Basics