Glossary
Each term means one thing in every chapter, and a chapter links its first use of a term here.
- agent
- A system in which a model directs its own steps: it chooses which tools to call, reads their results and decides what to do next, until the task is done or a limit stops it. Building effective agents
- agent loop
- The cycle that runs an agent: send the context to the model, run the tools it calls, add their results to the context, and repeat until the model answers without a tool call or a stop condition, such as a step limit, ends the run. Building effective agents
- approval
- A pause before a tool call runs, in which a person or a rule approves or denies it. The agent waits for the answer, and a denied call returns a denial to the model instead of running. Tool Approvals
- guardrail
- A check that runs beside the agent to catch unsafe or off-policy input, output or tool use. Guardrails work in layers, because no single one is enough, and they add to access controls rather than replace them. A practical guide to building agents
- harness
- The program around a model that turns it into a working agent: it runs the loop, calls the tools, keeps the session log and enforces the agent's limits. Pi, OpenCode, Claude Code and Codex are harnesses. Running Agent Harnesses Durably on Vercel
- lethal trifecta
- Three capabilities that are dangerous together: access to private data, exposure to untrusted content, and a way to communicate externally. An attacker can trick an agent that has all three into sending the private data to them. The lethal trifecta for AI agents: private data, untrusted content, and external communication
- MCP
- Model Context Protocol. An open standard for connecting AI applications to the systems where data lives, through one protocol instead of a custom integration for each source. Introducing the Model Context Protocol
- model gateway
- The layer between an agent and its model providers. It gives the agent one way to call many providers and handles shared concerns such as retries and prompt caching. Vercel's AI Gateway is a hosted example. AI Gateway FAQ
- prompt injection
- An attack that hides instructions in content the agent reads, such as a web page or an email, so the model follows them as if they came from the user. It works because trusted and untrusted text share one context. The lethal trifecta for AI agents: private data, untrusted content, and external communication
- sandbox
- An isolated environment, such as a microVM, where an agent runs code it wrote or code nobody has vetted, kept apart from the machine that runs the agent. Vercel Sandbox
- session log
- The ordered record of a session: each message, tool call and tool result as it happened. A harness can rebuild a run from it, so a turn that crashed or timed out resumes instead of starting over. Running Agent Harnesses Durably on Vercel
- workflow
- A fixed sequence of model calls and tools that the code lays out in advance. Unlike an agent, the model never chooses the next step. Building effective agents