AI Agent Glossary

The definitive reference for Claude Code operators. 40+ terms covering agent architecture, security, orchestration, and governance.

ABCDEFGHIJKLMNOPQRSTUVWXYZ

A

Agent

A Claude Code instance running autonomously with a defined identity, policy scope, and task. Every agent in Sentrely has a unique identity, audit trail, and permission set.

Agent Identity

A named, trackable persona assigned to each Claude agent. Enables per-agent RBAC policies, cost tracking, and audit attribution — you always know which agent did what.

Agent Policy

A YAML-defined set of rules governing what an agent can and cannot do. Policies specify allowed services, actions, resources, and budget limits — enforced at the gateway layer on every request.

Approval Gate

A checkpoint where an agent's requested action is paused until a human approves it via Slack or Telegram. Used for destructive operations, sensitive data access, and production deployments.

Audit Trail

The immutable, chronological record of every action an agent takes through the gateway. Logs every API call, tool invocation, file change, and approval decision with timestamps and agent identity.

Autonomous Agent

A Claude Code agent running with --dangerously-skip-permissions where it executes tool calls without interactive confirmation. Powerful but dangerous without a control plane.

A2A Messaging

Agent-to-Agent messaging — the protocol allowing Claude agents to send structured messages to each other through the gateway. All messages are logged and subject to policy enforcement.

B

Blast Radius

The maximum scope of damage an agent can cause if it malfunctions. Sentrely minimizes blast radius through least-privilege policies and resource-scoped permissions.

Budget Alert

A notification triggered when an agent's token usage approaches a defined threshold. Prevents cost surprises by giving operators time to act before limits are reached.

Butler

Sentrely's built-in Slack bot that handles operational notifications — approval requests, agent alerts, status updates, and budget warnings. All Butler actions are audited.

C

Circuit Breaker

A safety mechanism that automatically stops an agent after repeated failures. If an agent enters a runaway loop, the circuit breaker trips and kills the session before costs escalate.

Claude Code

Anthropic's agentic coding tool. Claude Code is the engine; Sentrely is the control plane that makes it safe to run Claude Code against real infrastructure in production.

Context Window

The maximum token count Claude can hold in working memory. Sentrely tracks per-session token usage to help operators detect when agents are hitting context limits.

Control Plane

The management layer sitting between Claude agents and the resources they access. Sentrely is a control plane — it governs what agents can do, logs everything, and provides the dashboard for oversight.

Cost Control

Mechanisms preventing agents from generating unexpected API bills. Includes per-agent token budgets, per-session limits, alert thresholds, and automatic session termination.

D

Deny-by-Default

Security posture where agents have zero permissions until explicitly granted. A new agent in Sentrely cannot access any service until an operator defines a policy granting specific permissions.

Destructive Operation

Any agent action that deletes, overwrites, or irreversibly modifies a resource. Sentrely can require human approval for these even when the agent has the underlying permission.

E

Escalation

The process by which an agent's action is routed to a human for review. Sentrely escalations arrive in Slack or Telegram with full context and a one-click approve/deny interface.

G

Gateway

The central proxy through which all agent traffic flows. Sentrely intercepts every request — API calls, git operations, cloud service access — and applies policy checks before forwarding.

GDPR

EU data privacy regulation applying when agents process personal data of EU residents. Sentrely's audit trails and policy enforcement help demonstrate GDPR compliance for agent workflows.

H

HIPAA

US healthcare privacy regulation applying when agents access Protected Health Information. Sentrely enables HIPAA-compliant operations through VPC deployment, encrypted audit trails, and access controls.

Human-in-the-Loop (HITL)

Operational model where certain agent actions require explicit human approval. Sentrely implements HITL through Slack/Telegram approval gates — configurable per operation type and risk level.

L

Least Privilege

Security principle: every agent gets only the minimum permissions required for its task. The opposite of running Claude Code with your personal AWS admin credentials.

Log Retention

Configured duration for keeping audit logs before archival. Sentrely supports configurable retention with automatic archival for compliance requirements (SOC 2: 1 year, HIPAA: 6 years).

M

Managed Control Plane

A hosted, maintained control plane used as a service. Sentrely is a managed control plane — you configure policies and agents; we handle enforcement, logging, dashboards, and upgrades.

Multi-Agent Orchestration

Coordination of multiple Claude agents working together. Sentrely supports fan-out (parallel subtasks), pipeline (sequential handoff), and supervisor patterns with audited handoffs.

P

Per-Agent Identity

Assigning each agent instance its own unique identity rather than sharing credentials. Enables granular audit trails, fine-grained policies, and precise cost attribution.

Pipeline Pattern

Multi-agent pattern where agents hand off work sequentially. Sentrely tracks pipeline stages in the audit trail and alerts if any stage stalls.

Policy-as-Code

Defining agent permissions in version-controlled YAML files rather than UI toggles. Sentrely policies go through code review, have change history, and can be tested before deployment.

R

RBAC

Role-Based Access Control. Sentrely's Casbin-powered RBAC lets operators define roles (code-reviewer, deployer, analyst) and assign agents to them rather than managing permissions individually.

Rate Limiting

Restricting how many requests an agent can make within a time window. Prevents agents from overwhelming external APIs or generating excessive costs through rapid-fire requests.

Runaway Loop

Failure mode where an agent gets stuck repeating the same action. Sentrely detects runaway patterns and triggers circuit breakers before costs escalate or damage occurs.

S

Session

A bounded instance of an agent's execution. Sentrely tracks sessions with start/end times, total token usage, actions performed, and final status — the primary unit of cost accounting.

Session Kill Switch

One-click mechanism in the dashboard to immediately terminate a running agent session. Sends termination signal, revokes credentials, and logs the action — all in under 10 seconds.

SOC 2

Compliance framework requiring demonstrated controls for data security and availability. Sentrely's immutable audit logs and RBAC enforcement provide the evidence trail SOC 2 auditors require.

T

Token Budget

Maximum API tokens an agent can consume in a session or time period. When a budget is hit, the gateway alerts, pauses, or terminates the session automatically depending on configuration.

Tool Call

An action a Claude agent takes using one of its tools — reading a file, running a command, calling an API. Every tool call passes through Sentrely for policy evaluation and audit logging.

V

VPC Deployment

Running Sentrely inside a Virtual Private Cloud so agent traffic never traverses the public internet. Required for organizations handling sensitive data under HIPAA, GDPR, or SOC 2.

Y

YOLO Mode

Claude Code's --dangerously-skip-permissions flag. Skips interactive approval prompts, enabling true agentic automation. Sentrely provides the control plane that makes YOLO mode production-safe.

Every term here is a feature

Audit trails, RBAC, approval gates, cost controls, session management — all built into Sentrely's managed control plane.