9. Process Structure Elements

(Normative)

Beyond per-activity attributes, the BPM/Agent Stack defines structural elements that govern the relationships between activities. All elements derive from BPMN 2.0.

9.1 Swimlanes (Responsibility Assignment)

Visual and semantic separation of responsibilities across actors — the structural expression of the RACI matrix. In agent architectures: which agent or human is responsible for each activity. Responsibility SHALL be explicit, not inferred by the LLM at runtime.

9.2 Milestones (Authorization Gates)

Named checkpoints that mark phase completion and optionally require authorization before proceeding. In agent architectures: governance checkpoints where human review, approval, or authorization is required. These are the operational expression of the Intent Stack’s governance interfaces.

9.3 Gateways (Decision Routing)

Explicit, structured branching and merging logic with defined semantics:

Gateway Type	BPMN 2.0 Semantics	Agent Mapping
Exclusive (XOR)	Exactly one path based on conditions	Deterministic routing — not LLM inference
Parallel (AND)	All paths execute simultaneously	Fan-out to multiple sub-agents with synchronization
Inclusive (OR)	One or more paths based on conditions	Selective parallel execution
Event-Based	Waits for one of several external events	Agent blocks until a trigger fires

The critical design decision: which gateway conditions SHALL be deterministic (Business Rule Tasks evaluated through DMN decision tables) versus which SHOULD involve LLM judgment. Deterministic conditions produce reproducible, auditable routing. LLM judgment produces adaptive, context-sensitive routing. The BPM/Agent Stack provides infrastructure for both; the choice is a governance decision at each gateway.

Gateways SHALL have typed semantics. An implementation SHALL NOT conflate gateway types — an exclusive gateway that sometimes behaves as inclusive violates the type system and undermines auditability.

9.4 Events (Exception Handling and Triggers)

A taxonomy of occurrences during process execution with defined handling semantics:

Event Type	BPMN 2.0 Scope	Agent Mapping
Timer	Scheduled triggers, timeouts	Timeout behavior, scheduled re-execution, periodic polling
Error	Failure conditions	Tool failure, API error, model refusal, resource exhaustion — each with typed handling
Escalation	Controlled elevation	Human-in-the-loop escalation when confidence is low or decision exceeds authorized scope
Signal	Broadcast notifications	Cross-agent broadcast
Message	Point-to-point communication	Structured inter-agent communication with typed payloads
Compensation	Rollback logic	Undo operations: delete created file, revert change, cancel API request

This taxonomy SHALL replace the blunt instruments in current frameworks (generic retry, maxIterations, timeout) with structured exception handling where each failure type has a defined response.

Error events SHALL have typed handling — a tool failure, an API timeout, and a model refusal are structurally different failure modes requiring different responses. Implementations SHALL NOT treat all errors as equivalent.

Escalation events SHALL be available for conditions exceeding agent authority. An agent that cannot resolve a situation within its authorized scope SHALL escalate rather than improvise.

9.5 Subprocesses (Governed Decomposition)

Nested process models with their own activities, gateways, events, and governance attributes, connected to the parent through a governed interface. In agent architectures: sub-agent delegation with defined inputs, expected outputs, boundary constraints, escalation triggers, and accountability.

A subprocess SHALL inherit all Boundary constraints from its parent process and MAY add additional constraints appropriate to its scope. A subprocess SHALL NOT relax any Boundary established by its parent. This is the process-level expression of Boundaries monotonicity.

9.6 Flows

Sequence Flows (BPMN 2.0) define explicit execution order within a participant’s scope, optionally with conditions. Message Flows define communication between separate participants with typed payloads. These SHOULD replace emergent execution ordering (where the LLM decides what happens next) with deterministic ordering where the process logic is known, while preserving agent judgment for genuinely adaptive situations.

Message flows between agents SHALL carry typed payloads conforming to the controlled vocabulary (Clause 10.1). Untyped, unstructured inter-agent communication SHOULD be treated as a governance gap.