DEFINITIONLast verified: April 2026

What Is an Agentic Swimlane? A Definitional Reference (2026)

An agentic swimlane is a BPMN 2.0 swimlane (the spec calls it a bpmn:lane) that represents an AI agent as the actor, inside a process flow diagram that includes one or more human lanes, one or more agent lanes, and optionally tool or system lanes. It answers the process question, in what order do the steps happen and who performs each one. It is distinct from the structural question, who reports to whom, which is the territory of agenticorgchart.com.

What goes in an agentic swimlane

A BPMN-correct agentic swimlane is built from a small set of elements, each with an entry in the OMG BPMN 2.0 specification.

Lanes (bpmn:lane, spec §10.5.5). One per actor type: human, agent, tool, system of record. Lanes belong to a pool.
Pools (bpmn:pool, spec §10.5.4). The participant grouping. Often the organisation or the process boundary.
Tasks with task-type markers. The agent step is normally a bpmn:serviceTask (spec §10.3.4), the human step a bpmn:userTask (spec §10.3.5), and a classifier-only step a bpmn:businessRuleTask (spec §10.3.6).
Gateways for decision points (bpmn:exclusiveGateway, spec §10.6.3 and bpmn:eventBasedGateway, spec §10.6.7).
Events at process boundaries and at signal points (bpmn:startEvent, bpmn:endEvent, bpmn:signalEvent).
Sequence flows within a pool, drawn solid (spec §10.7.1).
Message flows across pools, drawn dashed (spec §10.7.4). The canonical agent-to-agent handoff shape across organisational boundaries.

Shape 1: linear handoff (agent-augmented)

The most common starter shape. One human lane, one agent lane, one tool lane, with the work flowing left to right.

Linear-handoff swimlane. The human kicks the process off with a bpmn:userTask, the agent performs two bpmn:serviceTask steps, and the tool/system step writes to the system of record. End event terminates the process.Pattern source: synthesised from Anthropic, “Building Effective Agents” (Schluntz, December 2024) and the OMG BPMN 2.0 specification (object management group, 2014, latest formal/2014-01-03).

Shape 2: branching gateway (confidence-routed)

The agent classifies the inbound case, then a bpmn:exclusiveGateway routes high-confidence cases to auto-resolution and the rest to manual review by a human. The simplest pattern that captures “agent unless not sure, then human”.

Branching-gateway swimlane. Exclusive gateway diamond is the BPMN-correct shape; an outgoing path for each named branch.Source: shape per OMG BPMN 2.0 specification §10.6.3 (bpmn:exclusiveGateway). Pattern is observed in published Klarna customer service reports (2024) and in Anthropic case studies (2024 to 2025).

Shape 3: parallel processing (coordinator and workers)

One coordinator agent fans work out to two or more worker agents, each enriching a different facet, then merges. This is the orchestrator-workers pattern (Schluntz 2024, “Building Effective Agents”) drawn in swimlane form.

Parallel-processing swimlane. Each worker agent gets its own lane; parallel sequence flows from a fan-out task replace the deterministic parallel gateway when the workers are concurrently independent.Source: Anthropic, “Building Effective Agents”, anthropic.com/research/building-effective-agents, Schluntz, December 2024 (accessed April 2026). Shape conformance per OMG BPMN 2.0 §10.5.5 and §10.7.1.

Shape 4: escalation

The agent throws a bpmn:signalEvent when the case is ambiguous; a human catches the signal in another lane and resolves. The customer-support intake default. The escalation rate is the metric an operator cares about; Klarna's public reports name a target escalation envelope and a measured rate.

Escalation swimlane. Throwing signal event in the agent lane (filled triangle), catching signal event in the human lane (open triangle).Source: shape per OMG BPMN 2.0 §10.7.6. Pattern observed in Klarna AI assistant operator note, klarna.com (February 2024, accessed April 2026).

Shape 5: async loop (call-activity wrapped)

The agent's internal sense-think-act loop is non-deterministic and out of scope for BPMN's deterministic process semantics. The BPMN-correct pattern is to wrap the loop as a bpmn:callActivity (spec §10.4.4), so the wrapper process is BPMN-deterministic and the loop is detailed elsewhere.

Call-activity swimlane. Thicker-bordered task is the BPMN visual marker for a call activity. The agent loop is the called sub-process, defined and detailed separately.Source: OMG BPMN 2.0 §10.4.4 (bpmn:callActivity); workflow-vs-agent distinction per Anthropic, “Building Effective Agents”, Schluntz, December 2024.

What does NOT go in an agentic swimlane

Three things commonly appear on vendor BPMN diagrams that should not, and drawing them inline blows the diagram up. Each has a separate home.

The agent's internal sense-think-act loop. The loop is non-deterministic. Wrap it as a bpmn:callActivity; document the internals on the cornerstone reference (whatisanaiagent.com / how AI agents work).
The reporting structure. That is the org chart, not the process diagram. Cross-link to agenticorgchart.com for the structural view.
The deployment infrastructure. Network diagrams, container topology, and message broker layout belong on a system architecture diagram. Conflating them with the process flow defeats both views.

How an agentic swimlane differs from a flowchart

Flowcharts collapse actors into a single sequence. Swimlanes preserve the actor distinction explicitly. For agent-augmented processes, the actor distinction is exactly what the reader needs to see: which step is human, which is agent, where the handoff happens, and where the gate sits.

For the underlying definition of an AI agent, see whatisanaiagent.com. For the structural / org-chart view, see agenticorgchart.com. For the engineering reference, see buildingeffectiveagents.com.

Related pages on this site

Human vs agent swimlanes : the four-criterion decision rubric.
BPMN with AI agents : the spec-correct element reference.
Handoffs : agent-to-agent and agent-to-human patterns.
By process : five processes, one cited diagram each.