GLOSSARYLast verified: April 2026
AI Agent Workflow Glossary: Process Flow + BPMN Terms (2026)
A focused glossary covering the process-flow and BPMN-with-AI specific terms used across the site. Every BPMN element entry cites the OMG specification by section. Every framework primitive entry cites the relevant documentation page.
- A2A protocol#a2a-protocol
- Agent-to-Agent protocol. An open specification for cross-vendor agent handoffs over HTTP. Announced by Google and a coalition of partners in April 2025. Defines discovery and task delegation APIs.
- Agent lane#agent-lane
- A BPMN swimlane (bpmn:lane) in which the actor is an AI agent. Tasks within the lane are typically bpmn:serviceTask or bpmn:businessRuleTask.
- Arbiter pattern#arbiter-pattern
- A handoff pattern in which the agent attempts to resolve a case and escalates ambiguous cases to a human arbiter. The customer-support intake default.
- BPMN#bpmn
- Business Process Model and Notation. The OMG-standardised notation for process diagrams. Current normative version is BPMN 2.0.2 (formal/2014-01-03), adopted as ISO/IEC 19510:2013.
- BPMN call-activity#bpmn-call-activity
- A reusable sub-process invoked from a parent process. Used to wrap an AI agent's internal sense-think-act loop so the parent process stays BPMN-deterministic.
- BPMN end event#bpmn-end-event
- Process boundary event signalling termination of a sequence flow. Drawn as a thick-bordered circle.
- BPMN exclusive gateway#bpmn-exclusive-gateway
- A decision point with one outgoing path. The default gateway for branching. Drawn as a diamond with an X marker.
- BPMN lane#bpmn-lane
- A horizontal partition within a pool that groups activities by the actor performing them.
- BPMN message-flow#bpmn-message-flow
- Communication across pools. Drawn as a dashed arrow. The canonical agent-to-agent handoff shape across organisational boundaries.
- BPMN pool#bpmn-pool
- A participant grouping. Often the organisation, the process boundary, or an external participant.
- BPMN service-task#bpmn-service-task
- A task performed by an automated system or service. The canonical home for an autonomous AI agent step.
- BPMN signal-event#bpmn-signal-event
- A broadcast event used for escalations. Throwing variant in the agent lane, catching variant in the human lane.
- BPMN start event#bpmn-start-event
- Process boundary event signalling the start of a sequence flow. Drawn as a thin-bordered circle.
- BPMN user-task#bpmn-user-task
- A task performed by a human with software support. The canonical home for a human review step.
- Escalation#escalation
- The act of passing control from an agent to a human (or to a higher-tier agent) on detection of an out-of-scope, ambiguous, or high-stakes case.
- Handoff (agent-to-agent)#handoff-a2a
- A handoff between two agents within the same process. Documented in OpenAI Swarm and LangGraph as a primitive.
- Handoff (agent-to-human)#handoff-a2h
- A handoff from an agent to a human. Three sub-patterns: reviewer, arbiter, fallback.
- Human-in-the-loop#human-in-the-loop
- A process pattern in which a human reviews, approves, or modifies an agent's output before it is committed. Implemented as LangGraph's interrupt or as a downstream BPMN user-task.
- Interrupt (LangGraph)#interrupt-langgraph
- A LangGraph primitive that pauses graph execution at a designated node and waits for human input. The canonical human-in-the-loop primitive in the LangChain stack.
- Lane (swimlane)#lane-swimlane
- Synonym for bpmn:lane in conversational use. The horizontal partition within a pool.
- Reviewer pattern#reviewer-pattern
- A handoff pattern in which the agent acts and the human reviews the output before it is committed. LangGraph's interrupt is the canonical implementation.
- Sequence-flow#sequence-flow
- An arrow within a single pool indicating order of execution. Drawn solid, distinct from message-flow which is dashed.
- Swarm handoff (OpenAI)#swarm-handoff
- The handoff primitive in OpenAI Swarm. An agent returns a reference to another agent and the runtime continues under the receiving agent.
- Workflow vs agent (Anthropic)#workflow-vs-agent
- A distinction defined in Anthropic's December 2024 paper Building Effective Agents. A workflow is deterministic, controlled by predefined paths. An agent is non-deterministic, choosing its own steps within a tool budget. The distinction governs whether a step is BPMN-modellable as deterministic or wrapped as a call-activity.
Section A
Source: google.github.io / A2A (accessed April 2026)
Source: OMG BPMN 2.0 §10.5.5
Source: Klarna AI assistant operator note, February 2024
Section B
Source: omg.org / spec / BPMN / 2.0
Source: OMG BPMN 2.0 §10.4.4
Source: OMG BPMN 2.0 §10.7.2
Source: OMG BPMN 2.0 §10.6.3
Source: OMG BPMN 2.0 §10.5.5
Source: OMG BPMN 2.0 §10.7.4
Source: OMG BPMN 2.0 §10.5.4
Source: OMG BPMN 2.0 §10.3.4
Source: OMG BPMN 2.0 §10.7.6
Source: OMG BPMN 2.0 §10.7.1
Source: OMG BPMN 2.0 §10.3.5
Section E
Section H
Section I
Source: langchain-ai.github.io / langgraph (accessed April 2026)
Section L
Source: OMG BPMN 2.0 §10.5.5
Section R
Section S
Source: OMG BPMN 2.0 §10.7.1
Source: github.com / openai / swarm (accessed April 2026)
Section W
Source: Schluntz, anthropic.com / research / building-effective-agents (December 2024)