Reference
AI agent glossary
The vocabulary of production multi-agent systems — what each term means, how it behaves in a real runtime, and the mistake people usually make with it. 20 terms across 6 areas.
Execution
Correction nudge
A short corrective instruction the runtime injects into a run when a sub-agent's routing intent is ambiguous or looping — a soft, bounded retry instead of a hard failure.
ReAct loop
The reason-act-observe cycle an agent runs in — the model reasons about the task, calls a tool, reads the result, and repeats until it produces a final answer or hits a limit.
SSE streaming (agent runs)
Server-Sent Events delivery of an agent run's output as it is produced, with durable chunk writes so an interrupted run can be replayed or resumed rather than restarted.
Step budget
The cap on how many steps a run may take before it is stopped, alongside separate independent caps for transfers, delegations, and cross-agent hop depth.
Observability
Execution trace
The step-by-step record of what actually happened inside a run — every routing decision, tool call, and handoff — separate from the conversation history the user sees.
Termination reason
The explicit code recorded when a run ends, stating whether it completed normally or which specific limit, routing error, or failure stopped it.
Orchestration
Agent-to-agent (A2A) dispatch
Delegating work from one agent to a completely separate, externally-linked agent — across an organisational or data boundary — bounded by hop limits and cycle detection.
Delegation (agent handoff)
A handoff in which the caller asks another sub-agent for a result on a specific sub-task, receives it back, and stays in control of the conversation.
Multi-agent orchestration
Coordinating several specialised agents within one run — deciding which agent handles what, how work is handed between them, and how the result is assembled into a single answer.
Root agent
The entry point of a flow — it holds shared instructions and routes work to sub-agents, but never generates an answer itself.
Routing graph
The directed graph of permitted handoffs between sub-agents, declaring which agent may transfer or delegate to which other agents.
Sub-agent
A specialist agent inside a multi-agent flow, with its own model, instructions, tools, and skills, that produces the actual answers a run returns.
Transfer (agent handoff)
A handoff in which the target sub-agent becomes the new active responder for the rest of the conversation, and the original sub-agent stops.
Platform
Safety & Control
Fail-closed
A safety default where an unresolvable permission decision becomes a denial rather than an approval — so ambiguity blocks the action instead of allowing it.
Permission mode
The agent-level setting that decides whether tool calls run automatically, require human approval, or are blocked — with per-tool overrides taking precedence.
Tools & Integration
MCP server
A service that exposes tools, resources, or prompts over the Model Context Protocol, giving any compatible AI client access to its whole catalog through one connection.
Model Context Protocol (MCP)
An open protocol that lets AI applications connect to external tools and data sources through a standard interface, so a tool built once works with any MCP-compatible client.
Skill (agent skill)
A bundle of specialist instructions a sub-agent loads on demand mid-run, instead of carrying every domain's knowledge in one oversized system prompt.
Tool call
A model-initiated request to execute a named function with structured arguments, resolved against a permission policy before the platform runs it and returns the result.
These aren't abstractions — they're the runtime
Every term here describes behaviour Anter actually enforces: validated routing graphs, independent transfer and delegation budgets, fail-closed permissions, and a full execution trace for every run.