Definitions
Glossary
The vocabulary these loops assume, defined once so the loop pages do not have to re-explain it. Every definition cites a primary source.
- Agent harness An agent harness is the environment an AI agent runs inside, providing its tools, permissions, memory, and guardrails. Claude Code is a harness; the model alone is not.
- Agent loop An agent loop is a system that runs an AI agent on a heartbeat toward a verifiable goal, discovering work, doing it, checking it, and recording state between runs.
- Agent loop failure modes Agent loops fail in five recognisable ways: runaway cost, scope creep, a checker that rubber-stamps, a silent no-op, and a stopping condition the agent satisfies too easily.
- Agent skills Agent skills are installable instruction files that give an AI agent a reusable capability, packaging domain knowledge and process so it is not re-derived every session.
- Agent vs skill vs loop vs MCP An agent is the thing that acts, a skill teaches it one job, a loop is a job with a finish line, and an MCP is how it reaches an outside system. Four layers, often confused.
- Agentic marketing Agentic marketing is marketing work executed by AI agents that plan, act, and verify across multiple steps, with humans setting goals and approving output.
- Content provenance Content provenance in AI marketing is the disclosed record of who or what produced a piece of content, who approved it, and what sources back its claims.
- Context engineering, a term Anthropic uses for [managing what an agent knows at inference time](https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents), for marketing Context engineering for marketing is curating what an AI agent knows before it works, positioning, voice, proof, constraints, so output starts specific instead of generic.
- Human-in-the-loop marketing Human-in-the-loop marketing places explicit human approval gates inside agent-run marketing systems, so autonomy handles the work and humans keep the judgment.
- llms.txt llms.txt is a plain-text file at a site's root that gives AI agents a curated map of the site, what it offers, where the key pages are, and how to act on them.
- Maker-checker pattern The maker-checker pattern splits AI work between an agent that produces and a separate agent that verifies, because models grade their own work too generously.
- Marketing agent A marketing agent is an AI system that performs marketing work autonomously across multiple steps and tools, pursuing a goal rather than responding to single prompts.
- Marketing agent vs marketing automation Automation executes a fixed workflow you designed; an agent reads the current state and chooses its next action toward a goal. The difference is who decides the steps.
- Verifiable stopping condition A verifiable stopping condition is a finish line a machine can check, the difference between an agent task that ends when done and one that ends when it feels done.
- Vibe marketing Vibe marketing is directing AI systems to produce marketing output by describing intent and taste, then curating results, rather than executing the craft by hand.
- What not to automate Three kinds of marketing work should not be handed to an agent: decisions that set direction, claims that create liability, and anything where being wrong is expensive and hard to reverse.
These definitions describe what the loops do. Start with brain-build if you want the one that runs first.