An AI-native organization is still an organization: a set of business processes, designed deliberately, that produce consistent outcomes at scale. Most of the functions in those processes will not be performed by people. But the process must exist, must be explicit, and must be correct — or the organization is not AI-native, it is AI-assisted chaos.
Process correctness cannot be verified by running the organization and observing what breaks. It must be verified in simulation, before deployment, against the edge cases and failure modes that production will eventually surface. A world model — a formal, executable representation of the business process — is the only instrument that makes this possible. Without it, every structural decision is a guess. With it, the organization can test, compare, and improve process designs the way engineers test circuits: offline, systematically, before anything goes live.
47 structural patterns for AI-native business processes. Each one is a proven topology: the scenario it solves, the failure modes it introduces, and the value properties it carries.
Linear execution — one step completes before the next begins. Routing fixed at design time.
| ID | Pattern | Topology |
|---|---|---|
| 10.11 | Pipeline | Linear chain — each node transforms context before passing forward |
| 10.12 | Router | Classifier dispatches to one of N specialist branches |
| 10.13 | RAG Pipeline | Query encoder → retriever → reranker → generator with grounded context |
| 10.14 | Retry-Fallback | Primary agent with retry gate; fallback chain on repeated failure |
| 10.15 | Evaluator-Optimizer | Generator → Critic loop until acceptance or budget exhaustion |
| 10.16 | Staged Pipeline with Per-Stage Feedback | Sequential stages each with inline critic; back-edge repairs stage output before advancing |
One or more downstream branches activated based on conditions or first-complete semantics.
| ID | Pattern | Topology |
|---|---|---|
| 20.21 | Multi-Choice (OR-Split) | One or more of N branches activated based on runtime conditions |
| 20.22 | Multi-Merge | Each completing branch independently triggers downstream — no sync |
| 20.23 | Structured Discriminator | First of N parallel branches to complete triggers outgoing flow |
| 20.24 | Blocking Discriminator | First wins; blocks re-firing until all branches complete |
| 20.25 | Cancelling Discriminator | First wins; remaining branches cancelled immediately |
| 20.26 | Structured Partial Join | Wait for M-of-N branches; remaining complete normally |
| 20.27 | Blocking Partial Join | Wait for M-of-N; blocks re-entry until all N complete |
| 20.28 | Generalised AND-Join | Wait for all currently active branches (count determined at runtime) |
| 20.29 | Deferred Choice | Branch chosen by external event, not internal condition; choice deferred to environment |
| 20.30 | Interleaved Parallel Routing | All tasks execute exactly once, one at a time, in any order |
| 20.31 | Milestone | Task enabled only if process is in specific state; missed milestone disables task |
Multiple work items or branches executing concurrently.
| ID | Pattern | Topology |
|---|---|---|
| 30.31 | Orchestrator-Workers | Central agent decomposes task, dispatches workers, synthesizes results |
| 30.32 | Competitive Evaluation | N generators compete in parallel; Judge selects the best candidate |
| 30.33 | Multi-Tier Orchestration | 3-level hierarchy: root orchestrator → domain orchestrators → leaf agents |
| 30.34 | MI Without Synchronization | Concurrent instances; process continues without waiting for all |
| 30.35 | MI with Design-Time Count | Fixed N instances at design time; all must complete before continuing |
| 30.36 | MI with Runtime Count | Count determined at runtime before start; all must complete |
| 30.37 | MI Without Prior Runtime Knowledge | Count unknown; new instances created dynamically; runtime completion condition |
| 30.38 | Static Partial Join for MI | Fixed N-of-M MI threshold; remaining instances complete normally |
| 30.39 | Cancelling Partial Join for MI | N-of-M MI threshold; remaining instances cancelled |
| 30.40 | Dynamic Partial Join for MI | Dynamic instance creation; runtime completion condition |
| 30.41 | Thread Merge | N execution threads within one process instance collapsed to one |
| 30.42 | Thread Split | One branch creates N new execution threads within same instance |
Repeated execution — loops, recursion, tool-use cycles.
| ID | Pattern | Topology |
|---|---|---|
| 40.41 | Tool-Use Loop (ReAct) | Reasoner selects tool, observes result, iterates until task complete |
| 40.42 | Structured Loop | Pre/post-test loop with single entry and exit point |
| 40.43 | Recursion | Task invokes itself; hierarchical decomposition to base case |
Multiple agents or branches that must align — synchronizing merges, human review, mutual exclusion.
| ID | Pattern | Topology |
|---|---|---|
| 50.51 | Human-in-the-Loop | Confidence gate routes low-certainty outputs to human review queue |
| 50.52 | Consensus (M-of-N) | N independent validators must agree (M-of-N) before output is accepted |
| 50.53 | Local Synchronizing Merge | Sync only branches activated by most recent split; uses local routing data |
| 50.54 | General Synchronizing Merge | Fires when all active branches done or no more tokens can arrive |
| 50.55 | Critical Section | Mutually exclusive regions; only one executes at a time |
| 50.56 | Interleaved Routing | Tasks execute exactly once, one at a time, any order, runtime-determined |
Stopping work that is no longer needed — task, region, instance, or multi-instance scope.
| ID | Pattern | Topology |
|---|---|---|
| 60.61 | Cancel Activity | Single running task cancelled; work discarded |
| 60.62 | Cancel Case | Entire process instance terminated; all active tasks stopped |
| 60.63 | Cancel Region | Named set of tasks cancelled together by a trigger signal |
| 60.64 | Cancel MI Activity | All incomplete instances of a MI task cancelled; completed ones kept |
| 60.65 | Complete MI Activity | MI task force-completed; remaining instances withdrawn |
| 60.66 | Explicit Termination | Designated end node cancels all remaining active branches immediately |
External signals that start or resume work — transient vs. persistent delivery semantics.
| ID | Pattern | Topology |
|---|---|---|
| 70.71 | Transient Trigger | Signal lost if no task waiting; fire-and-forget |
| 70.72 | Persistent Trigger | Signal stored until consumed; no trigger is dropped |
Patterns that operate across multiple executions. Change the system, not just the output.
| ID | Pattern | Topology |
|---|---|---|
| 80.81 | Feedback Loop | Outcomes update model weights / prompts / retrieval stores |