Glossary
AI agent glossary.
The vocabulary that comes up once you move from reading about agents to running them. Every entry is one sentence you could quote, then a short explanation including the part people usually get wrong.
18 terms
- Agent babysitting
- Agent babysitting is the recurring operational work an always-on agent creates: restarts, key rotation, spend checks, log reading, and host patching.
- Agent environment
- An agent environment is the isolated place an agent runs: its own filesystem, process space, credentials, and resource limits, separate from other agents.
- Agent runtime
- An agent runtime is the software harness that runs an agent: it holds the loop, manages context, and exposes the tools the agent is allowed to call.
- Agent template
- An agent template is a pre-built agent definition with a job, a schedule, and a set of tools already configured, so deploying it produces a working agent immediately.
- AI agent
- An AI agent is a program that uses a language model to decide which steps to take toward a goal, and can call tools to carry those steps out.
- Always-on agent
- An always-on agent is one that stays running continuously so it can act on schedules and triggers without a person starting it.
- Approval mode
- Approval mode is a setting where an agent pauses before a specified action and waits for a human to sign off before carrying it out.
- BYOK (bring your own key)
- BYOK means connecting your own model provider API key to a platform, so model usage is billed to you by the provider instead of through the platform.
- Managed agent hosting
- Managed agent hosting is a service that provisions and operates the environments agents run in, so the customer configures agents rather than servers.
- MCP server
- An MCP server is a program that exposes tools, data, or prompts to an AI agent over the Model Context Protocol, a shared standard for connecting agents to external systems.
- Model drift
- Model drift is when an agent's output quality changes without its configuration changing, because the underlying model, an API, or a data source changed instead.
- Multi-agent team
- A multi-agent team is a group of agents that run together and communicate, each handling part of a task rather than one agent handling all of it.
- Secret vault
- A secret vault is encrypted storage for the credentials an agent needs, injected at runtime so keys are never written into prompts, code, or configuration files.
- Self-hosting
- Self-hosting means running an agent on infrastructure you own and operate, such as your own machine or a VPS, rather than on a platform that operates it for you.
- Silent failure
- A silent failure is when an agent stops working without producing any error, so its absence looks identical to a quiet day with nothing to report.
- Spend cap
- A spend cap is a limit on what an agent can spend on model usage, and a useful one halts execution at the limit rather than sending a notification.
- Trigger
- A trigger is an event that causes an agent run to start, such as an incoming message, an inbound webhook, or a scheduled time being reached.
- Unattended execution
- Unattended execution means an agent runs on a schedule or trigger with nobody present to watch it, approve its steps, or intervene.