Evergreen explainers for the building blocks of AI engineering.
LLM-driven program that decides which tools to call in a loop to accomplish a goal, rather than answering a single prompt. Modern coding IDEs all build on agent loops.
Terminal-based pair programmer. Uses git to track AI-generated edits; works with Claude, GPT-4, and local models.
Direct REST API for Claude models. Supports prompt caching, batch mode, structured outputs, and tool use.
Anthropic's Python and TypeScript SDK for building agentic apps on top of Claude. Wraps tool use, sessions, and the agent loop.
Anthropic's command-line coding agent. Runs in your terminal, uses Claude as the model, supports skills, hooks, slash commands, and MCP servers. Designed for vibe coding longer sessions with file context and tool use.
Open-source autonomous coding agent for VS Code. Multi-step task execution with human-in-the-loop confirmation.
OpenAI's coding-focused agent / CLI. Newer iteration of the Codex brand built for autonomous coding tasks with GPT-class models.
Discipline of curating exactly the right files, instructions, and history that go into an LLM's context window. Distinct from prompt engineering — the focus is on what the model sees, not how you ask.
Open-source autopilot for VS Code and JetBrains. Brings agent workflows, custom commands, and context providers into the editor.
AI-first IDE forked from VS Code. Integrates Claude, GPT, and Gemini for inline edits, multi-file refactors, and agentic workflows. Core daily driver for vibe coders.
Google's flagship model family and matching CLI. Strong at long-context analysis and structured output; Gemini Flash is the cheap workhorse for high-volume pipelines.
GitHub's in-editor AI assistant. Inline autocompletion and chat, now with agent mode and MCP support. Bundled with most paid GitHub plans.
Claude Code lifecycle interception. Hooks run before or after tool calls, prompts, or session events — useful for guardrails, logging, and policy.
Popular framework for chaining LLM calls, retrieval, and tools. Big surface area, occasionally controversial design, but the de-facto starter for many teams.
Framework specialised for retrieval and indexing — better defaults than LangChain when your app is mostly RAG.
Open protocol from Anthropic for connecting LLM clients to data sources and tools via standardised servers. Cursor, Claude Code, and Copilot all support MCP servers.
OpenAI's REST API for GPT, embeddings, image, and audio models. The widest tool-use ecosystem and the most third-party integrations.
Unified API for switching between LLMs from Anthropic, OpenAI, Google, Mistral, and dozens more. Useful for avoiding lock-in and price arbitrage.
API feature that lets you reuse a long static prompt prefix across many requests for a fraction of the token cost. Anthropic, OpenAI, and Google all offer it with slightly different rules.
Pattern where an LLM is given relevant documents at inference time (retrieved from a vector store or other index) rather than relying on training data alone. Standard recipe for grounding answers in your own data.
Claude Code abstraction: a reusable bundle of instructions, scripts, and context that the agent can load on demand. Like a slash command with structured behaviour.
Custom user-defined command in Claude Code (and similar tools) invoked with /name. Used to encapsulate reusable workflows like /review or /cache.
Spawned child agent with its own context and tool set, called from a parent agent. Used to parallelise work and isolate context.
Practice of reducing the number of tokens sent to and received from an LLM without losing task quality. Includes prompt caching, message pruning, context window management, and structured output.
Working style where you steer an AI agent through a coding task in natural language without writing the code yourself line by line. Popularised by Andrej Karpathy in early 2025.
AI-native IDE from Codeium with Cascade agent for cross-file edits. A direct alternative to Cursor with a different agent loop.