Agent Loop — The cycle where an AI agent reads context, decides on an action, executes it, observes the result, and repeats until the task is complete or it needs human input.
Agent Mode — A feature in AI coding tools (Cursor Agent, Copilot Workspace, Claude Code) where the AI acts autonomously across multiple files, planning and executing multi-step changes.
A2A (Agent-to-Agent) — Google's protocol for agents to discover and communicate with each other. Complements MCP.
Agentic Coding — The practice of using AI agents (not just assistants) to write, test, debug, and deploy software autonomously within defined boundaries.
Background Agent — An AI agent that runs long tasks autonomously (minutes to hours) without requiring the user to watch. Claude Code supports this.
Batch API — An API mode that processes multiple requests asynchronously at a discount (typically 50% off). Used for non-real-time workloads like content processing.
Chain-of-Thought (CoT) — A prompting technique where the model is asked to show its reasoning step by step before giving a final answer. Improves accuracy on complex tasks.
CLAUDE.md — A project configuration file for Claude Code. Contains commands, code style, architecture notes, and gotchas. Similar to .cursorrules for Cursor.
Context Window — The maximum amount of text (measured in tokens) that a model can process at once. Larger windows allow the agent to "see" more of your project.
Few-Shot Prompting — Including 2-5 examples in the prompt to calibrate the model's output format and style. More reliable than zero-shot for structured tasks.
Fine-tuning — Additional training applied to a base model on task-specific data to shift its behavior. Less common for agentic coding than retrieval + prompt engineering; expensive to iterate.
Function Calling — The ability of an LLM to output structured requests to call external functions/tools. The foundation of agent tool use.
Grounding — Anchoring the model's output in verifiable context (retrieved documents, tool-returned data, file contents) rather than parametric memory. Reduces hallucinations.
Guardrails — Programmatic checks wrapped around agent input/output that block policy violations (PII leaks, unsafe commands, off-scope topics). OpenAI Agents SDK exposes this as a primitive; other frameworks implement via hooks.
Hallucination — When an AI model generates information that appears correct but is fabricated. A key risk in agentic coding that requires verification.
ISR (Incremental Static Regeneration) — A Next.js feature that regenerates static pages on demand or at intervals, combining static performance with dynamic freshness.
MCP (Model Context Protocol) — An open standard by Anthropic for connecting AI agents to external tools and data sources. Think USB for AI.
MCP Server — A program that exposes tools and resources to AI agents via the MCP protocol. Can run locally (stdio) or remotely (HTTP).
Multi-Agent — An architecture where multiple AI agents collaborate on a task, each with a specialized role. Frameworks: CrewAI, AutoGen, LangGraph.
Prompt Caching — Storing frequently reused prompt segments (system prompts, few-shot examples) so subsequent requests cost less. Anthropic charges 0.1x for cached tokens.
Prompt Engineering — The practice of crafting effective instructions for AI models. In agentic coding, this includes system prompts, CLAUDE.md files, and tool schemas.
Prompt Injection — An attack where untrusted text (fetched webpage, file content, tool output) contains instructions the model obeys as if they came from the user. Mitigated by sandboxing, output filtering, and treating all tool-returned text as data, not instructions.
RAG (Retrieval-Augmented Generation) — A pattern where relevant documents are retrieved from a database and included in the prompt to ground the model's response in factual data.
Sandboxing — OS- or container-level isolation that restricts an agent's filesystem, network, and process access. Codex CLI sandboxes by default; Claude Code offers /sandbox for opt-in isolation. Essential for autonomous runs.
Server Components (RSC) — React components that render on the server. In Next.js, they can directly query databases and call APIs without client-side JavaScript.
Structured Output — Forcing an LLM to return data in a specific format (JSON, XML) using tool schemas. More reliable than asking for raw JSON.
Tool Use — The ability of an AI agent to call external functions — reading files, running commands, querying databases, calling APIs. The defining capability of agentic systems.
Tokens — The basic units of text that LLMs process. Roughly: 1 token = 0.75 words. Pricing is per million tokens (MTok).
Vibe Coding — A workflow where the developer describes intent at a high level and the AI agent handles implementation details. Coined by Andrej Karpathy.
Search for a command to run...