The Vibe Father
Reference
Agentic coding invented a decade's worth of vocabulary in about two years, and most explanations of it assume you already know the other half of the vocabulary. This is the fix: every term you'll actually encounter, defined in plain language, with the honest caveats included. Bookmark it; we keep it current.
Agent
A language model wrapped in a loop that can read files, run commands, observe the results, and iterate toward a goal. The loop is the definition — a model that only completes text is a chatbot, however good the text. Full breakdown in What Is an AI Coding Agent?
Harness
Everything around the model: the tool definitions, permission system, context management, prompts, and orchestration that turn a raw model into a usable agent. Two products on the identical model can perform wildly differently because the harness differs — which is why we argue the harness matters more than the model.
Model vs. CLI
The model is the neural network (Claude, GPT, Gemini); the CLI is the terminal program that drives it (Claude Code, Codex CLI, Aider). People say "I use Claude" when they mean both. The distinction matters because some CLIs let you swap models freely and some lock you to one lab.
Context window
The model's working memory — the maximum amount of text (your prompt, the files it read, the conversation so far) it can consider at once. When it fills, old material falls out or gets summarized, and the agent starts forgetting decisions it made an hour ago. Every long-session pathology traces back here.
Token
The unit models read and bill in — roughly three-quarters of an English word, less for code. Context windows are measured in tokens; API pricing is per million tokens. When a tool feels expensive or forgetful, tokens are the reason.
System prompt
The standing instructions a harness sends before your message: the agent's role, rules, tool descriptions, and safety constraints. You mostly can't see it, but it shapes behavior more than anything you type. Project files like CLAUDE.md or AGENTS.md are your writable extension of it.
BYOK (Bring Your Own Key)
Using your own API key from a lab instead of paying a middleman's bundled subscription. You pay the lab's raw per-token rates and skip the markup, at the cost of a little setup. Setup takes about ten minutes per lab — see our BYOK setup guide and why we bother.
MCP (Model Context Protocol)
The open standard, released by Anthropic in late 2024, for connecting agents to external tools — databases, browsers, GitHub, anything. Write one MCP server and every MCP-compatible agent can use it. The USB-C of agent integrations; the long version is in our MCP explainer.
Multi-agent / orchestration
Running several agents in parallel or in sequence on one project — splitting work, reviewing each other, exploring alternatives. Genuinely powerful and genuinely easy to turn into chaos; the coordination layer is the hard part. Our field guide to multi-agent teams covers what actually works.
Coordinator / Builder / Scout / Reviewer
The common role split in multi-agent setups. The coordinator plans and delegates; builders write code; scouts research the codebase or docs before anyone builds; reviewers check the work. Roles are just prompts plus permissions — but the separation of author from reviewer is the part with real teeth. See which model fits which role.
Verification gate
A checkpoint an agent's work must pass before it counts as done — tests green, build clean, a reviewer's approval. The antidote to agents that confidently declare victory on broken code. The rule: the gate runs the checks itself and never trusts the author's claim.
Checkpoint
A commit made purely so you can get back to this exact state — typically right before letting an agent run autonomously. Costs seconds, converts any agent disaster into a thirty-second rollback. The habit and the commands are in Checkpoints & Git Worktrees.
Git worktree
A git feature that gives you multiple working directories from one repository, each on its own branch. The clean way to run parallel agents: each gets its own directory, nobody overwrites anybody, you merge the winners. Same guide as above.
Session limit
The cap — hourly, per-session, or weekly — that subscription coding tools impose on usage. It always arrives mid-refactor. Surviving it is a learnable skill: handoff files, deliberate compaction, work on disk instead of in chat. Full playbook in the session limit survival guide.
Compaction
Summarizing a long conversation so it fits back in the context window. Auto-compaction fires at the worst moment and a summarizer decides what to keep; deliberate compaction at a green checkpoint keeps the facts you chose. Prefer the second.
Hallucination
The model stating something false with full confidence — an API that doesn't exist, a package name it invented, a test result it never saw. Not a bug being fixed next quarter; a standing property of the technology that your workflow has to price in. Hence verification gates.
SWE-bench Verified
The most-cited agentic coding benchmark: real GitHub issues from real Python repos, scored by whether the agent's patch passes the repo's own tests, on a human-validated subset of 500 problems. Useful signal, gameable like all benchmarks — our SWE-bench explainer covers how to read the numbers.
Terminal-Bench
A benchmark for the messier reality: full tasks in a live terminal — installing, configuring, debugging, operating a system — not just producing a patch. Closer to what daily agent use actually feels like.
LiveCodeBench
A contamination-resistant coding benchmark that continuously adds problems published after models were trained, so nobody can ace it by having memorized the test set. The score to check when you suspect a model is benchmark-tuned.
Vibe coding
Building software by directing agents in natural language and judging results by behavior rather than reading every line. Ranges from "prototyping superpower" to "shipping code nobody understands," depending entirely on the guardrails. Our honest guide draws that line carefully.
Handoff file
A file — conventionally HANDOFF.md — an agent writes before a session ends: the goal, decisions made and rejected, current state, numbered next steps, gotchas. It's how work survives session death, because a file in the repo outlives any chat. The ritual is step one of the survival guide.
Terms we should add? Tell us. The field moves fast enough that this page is a living document, and the definitions above are the ones we'd bet our own projects on — because we do, daily.