AI Agent Glossary
Plain-English definitions of 40+ terms covering AI agents, safety, approvals, and building software with AI.
A
Agent #
An AI model running with a goal, a set of tools, and some autonomy to decide its next step. A coding agent like Claude Code reads files, runs commands, and edits code on its own until the task is done.
Agent Identity #
A named, trackable identity assigned to each AI agent instead of a shared login. It lets you scope permissions per agent, attribute costs, and know which agent did what.
Agent Policy #
A written set of rules for what an agent can and cannot do: which services, actions, and resources it may touch, and what spending limits apply. Policies are most useful when they are enforced outside the agent, not just described in its prompt.
AI Employee #
An AI agent given an ongoing role, such as building and maintaining a piece of software, rather than a one-off prompt. At Sentrely, AI employees are deployed and supervised by our engineers.
Approval Gate #
A checkpoint where an agent's requested action is paused until a human approves it, often through Slack. Used for destructive operations, sensitive data access, and production deployments.
Audit Trail #
A tamper-resistant, chronological record of what an agent did: API calls, tool invocations, file changes, and approval decisions, each with a timestamp and the agent's identity.
Autonomous Agent #
An agent that executes tool calls without asking for confirmation each time, for example Claude Code with --dangerously-skip-permissions. Fast, but it needs outside guardrails to be safe.
A2A Messaging #
Agent-to-agent messaging: structured messages one AI agent sends another to hand off or coordinate work. In production these messages should be logged like any other action.
B
Blast Radius #
The maximum damage an agent can cause if it malfunctions. Least-privilege permissions, scoped credentials, and backups you can rewind all shrink the blast radius.
Budget Alert #
A notification triggered when an agent's token usage or spend approaches a set threshold, giving a human time to act before the limit is reached.
C
Circuit Breaker #
A safety mechanism that automatically stops an agent after repeated failures, so a runaway loop ends before costs or damage pile up.
Claude Code #
Anthropic's agentic coding tool. It reads a codebase, runs commands, and makes changes on its own. It is the engine; production use still needs reviews, permissions, and backups around it.
Context Window #
The maximum number of tokens a model can hold in working memory at once. Long sessions that hit the limit start forgetting earlier decisions, which is why written notes and handoff docs matter.
Control Plane #
In AI agent setups, the management layer between agents and the resources they touch. It decides what agents may do, logs what they did, and routes risky actions to humans.
Cost Control #
Mechanisms that prevent agents from running up unexpected bills: per-agent token budgets, per-session limits, alert thresholds, and automatic session termination.
D
Deny-by-Default #
A security posture where an agent has no permissions until they are explicitly granted. The safe starting point for anything that touches production.
Destructive Operation #
Any action that deletes, overwrites, or irreversibly changes a resource. These are the actions most worth gating behind human approval, even when the agent technically has permission.
E
Escalation #
Routing an agent's action to a human for review, with enough context for them to approve or deny it quickly.
G
Gateway #
A central proxy that agent traffic flows through, so requests can be checked against policy, logged, or rate limited before they reach the outside world.
GDPR #
The EU data privacy regulation. It applies when software or agents process personal data of people in the EU, and it requires you to know where that data goes and who can access it.
H
HIPAA #
The US healthcare privacy regulation. It applies when software or agents handle Protected Health Information, and it requires access controls, encryption, and audit logs.
Human-in-the-Loop (HITL) #
An operating model where certain AI actions require explicit human approval before they happen. At Sentrely, important changes get sign-off from one of our engineers before they ship.
L
Least Privilege #
The security principle that every agent gets only the minimum permissions its task needs. The opposite of running an agent with your personal admin credentials.
Log Retention #
How long audit logs are kept before archival or deletion. Compliance frameworks set minimums, for example one year is a common SOC 2 expectation.
M
Managed Control Plane #
A control plane run for you as a service rather than built and operated in-house.
MCP (Model Context Protocol) #
The open protocol that lets agents like Claude Code connect to external tools and data sources through MCP servers. Powerful and risky: an over-permissioned MCP server can expose sensitive data or allow actions nobody intended.
MCP Server #
A connector exposing a tool or data source (GitHub, a database, AWS) to an agent over MCP. By default there is no permission layer between the agent and the server, so scope each server to least privilege.
Multi-Agent Orchestration #
Coordinating multiple AI agents working together, for example fan-out (parallel subtasks), pipeline (sequential handoff), and supervisor patterns.
P
Per-Agent Identity #
Giving each agent instance its own identity rather than shared credentials. Enables precise audit trails, fine-grained permissions, and cost attribution.
Pipeline Pattern #
A multi-agent pattern where agents hand off work sequentially, each stage taking the previous stage's output as input.
Policy-as-Code #
Defining agent permissions in version-controlled files rather than UI toggles, so changes go through review, have history, and can be tested.
R
RBAC #
Role-Based Access Control: permissions are attached to roles (reviewer, deployer, analyst) and identities are assigned roles, rather than managing permissions one by one.
Rate Limiting #
Restricting how many requests an agent can make in a time window, so it cannot overwhelm external APIs or run up costs with rapid-fire requests.
Rewind #
Restoring data and files to an earlier point in time after a bad change. Sentrely keeps a 35-day rewind on your data and files.
Runaway Loop #
A failure mode where an agent gets stuck repeating the same action, burning tokens and sometimes making the same bad change over and over.
S
Safety Check #
An automated test or scan every change must pass before it ships, such as tests, type checks, and security scans. Sentrely runs automated safety checks on every change.
Session #
A bounded run of an agent, with a start and end, total token usage, actions performed, and a final status.
Session Kill Switch #
A way to immediately stop a running agent session, revoke its credentials, and record that it happened.
SOC 2 #
A compliance framework that requires demonstrated controls for data security and availability, including evidence of who changed what and when.
T
Token Budget #
The maximum number of model tokens an agent may consume in a session or time period, after which it is alerted, paused, or stopped.
Tool Call #
An action an agent takes using one of its tools: reading a file, running a command, calling an API.
V
Vibe Coding #
Building software by describing it to an AI app builder (Base44, Replit, Lovable) and accepting what it generates. Great for mock-ups; production software also needs reviews, backups, and someone who answers when it breaks.
VPC Deployment #
Running software inside a Virtual Private Cloud so traffic stays off the public internet. Common for organizations handling sensitive data.
Y
YOLO Mode #
Claude Code's --dangerously-skip-permissions flag. Skips interactive approval prompts so the agent can work unattended. Safe only with outside guardrails: scoped credentials, reviews, and backups.
Rather not manage any of this yourself?
Sentrely is a managed AI team, backed by real engineers, that builds, ships and monitors production-grade software in accounts you own. Safety checks, human sign-off and rewind are built in.