AI Agent Workflow Diagrams: BPMN Swimlanes with Cited Examples (2026)
How agentic processes flow across teams. Visualised in BPMN 2.0, with a named public source for every swimlane on the site.
Linear handoff, branching gateway, parallel processing, escalation, async loop. Shape semantics per OMG BPMN 2.0 specification. Process examples drawn from Klarna, Cisco/ServiceNow, SAP/Concur, Anthropic/Clay, Cognition Devin.
The five canonical swimlane shapes
Almost every AI agent workflow diagram in the wild reduces to one of five shapes. Each is named, illustrated in BPMN-correct form, and linked to its deep-dive page. Each shape has at least one cited public source.
Agent-augmented swimlane
One human lane, one agent lane, one tool lane. A request is received, classified by the agent, and persisted by a system step. The most common starter shape.
Confidence-routed swimlane
Agent classifies, then a bpmn:exclusiveGateway routes high-confidence cases to auto-resolution and the rest to human review.
Coordinator and workers
One coordinator agent fans work out to two or more worker agents, each enriching a different facet, then merges the result. The orchestrator-workers pattern in swimlane form.
Escalation swimlane
Agent throws a bpmn:signalEvent when ambiguous; a human catches it in another lane and resolves. The customer-support intake default.
Call-activity wrapped loop
The agent's internal sense-think-act loop is wrapped as a bpmn:callActivity. The wrapper process is BPMN-deterministic; the loop is detailed elsewhere.
Cross-pool message flow
Agent-to-agent handoff across organisational boundaries is a bpmn:messageFlow, not a sequence flow. Documented in OpenAI Swarm and LangGraph.
The single most-asked operator question
Should this step go in the human lane or the AI agent lane? Vendor pages duck the question because their product is sold as agent-everywhere.
The neutral answer is a four-criterion rubric: autonomy required, reversibility of the action, regulatory constraint, and cost of error. Each criterion routes the step to a different placement.
Read the decision rubric →Does the step need genuine decision-making?
Rule-following steps belong in an agent lane. Decision-under-uncertainty steps benefit from human review.
Can the action be undone?
Sending money or filing a regulatory return: irreversible. Drafting a reply: reversible. Irreversible steps need a human gate.
Is human review required?
Some sectors (finance, healthcare, legal) require an identifiable human in the chain regardless of confidence. Cite the framework.
What is the downside of being wrong?
If the cost of a wrong action exceeds the cost of human review, review is the cheaper path even at high model confidence.
BPMN 2.0 spec correctness is the editorial differentiator
Most pages on the open web that show an AI agent “BPMN” diagram render an autonomous agent step as a generic bpmn:task rather than the specification-correct bpmn:serviceTask or bpmn:businessRuleTask. The distinction matters: BPMN 2.0 reserves task-type markers for exactly this kind of automation classification, and tool round-tripping (Camunda, bpmn.io, Signavio) depends on it.
The reference page walks each relevant element with its agent-process meaning and cites the OMG specification by section.
Read the BPMN element reference →By process, with a cited example each
Across processes, the same BPMN shapes recur. The differentiator is which steps are agent-driven and where the human gates sit. Each process page carries one BPMN-correct swimlane and one named public case study.
Customer support intake
Tier-1 agent + escalation to human via signal event.
Invoice processing
Agent-only swimlane with mandatory bpmn:userTask approval gate before payment.
IT ticketing
Agent-driven triage and routing, with human-as-arbiter for ambiguous cases.
Sales pipeline
Lead enrichment, qualification, and outbound personalisation by agent, with human approval before send.
Engineering task flow
Agent-driven loop wrapped in a deterministic process with human review on PR before merge.
All processes →
Index page covering all five process pages with the canonical topology pattern for each.
Handoffs: agent-to-agent and agent-to-human
Handoff primitives are documented in OpenAI Swarm and LangGraph but consolidated nowhere. The reference page walks the patterns, the BPMN representation (message-flow vs sequence-flow vs signal event), and the common failure modes.
Handoff patterns →How to draw your own swimlane
draw.io, Lucidchart, Miro, Whimsical, bpmn.io, Camunda Modeler, mermaid and react-flow. Compared on BPMN 2.0 fidelity, bundle size, accessibility, build-time vs runtime, SEO, and licensing. No rankings, no “best”.
Compare the tooling →In this cluster
Four sites cover the AI-agent-concepts territory together. Each answers a different operator question.
Methodology and corrections
The site is a reference, not a vendor pitch. Two citation disciplines run in parallel: every swimlane references a real, public, named source, and every BPMN-shaped diagram is checked against the OMG BPMN 2.0 specification. Where vendors are named, they appear as examples of patterns described, not as endorsements. There is no affiliate relationship with any AI vendor, BPMN tool vendor, framework maintainer, or cloud provider.
For more on how the site is built, the source pool, the editorial exclusions, and the update cadence, read the methodology page. For frequently asked questions, see the FAQ. For terminology, the glossary covers process-flow and BPMN-with-AI specific terms.