Agent orchestration patterns: sequential, parallel and router
Nova3AI editorial · Updated 11 September 2026 · Educational examples, not customer results
Choose the flow from the work's dependencies. Sequential steps consume earlier results. Parallel branches handle independent work. A router selects a path based on the request. One system may combine these patterns.
Compare the first three patterns
| Pattern | Use it when | Test before relying on it |
|---|---|---|
| Sequential | Lookup needs a validated reference; drafting needs lookup results | A failed earlier step must prevent an unsupported answer |
| Parallel | Independent checks can run against the same input | A missing branch, timeout or conflict must have a defined merge policy |
| Router | Orders, billing and technical requests need different handlers | Mixed, unknown and ambiguous requests need a fallback |
A support router with sequential checks
For an order query, validate the reference before calling a read-only tool. For a refund, route to a person. The router decides where the request goes; it does not grant the chosen specialist additional permissions.
Do not parallelise dependent work
Drafting an order answer before the status lookup completes risks unsupported claims. Independent format and policy checks can run together, but the final response must wait for the required results. Define whether any failed check blocks release.
Bound latency and cost
Sequential latency accumulates across steps. Parallel execution can shorten waiting time for independent branches, but still consumes the work of every branch and adds coordination overhead. Measure the actual critical path, tool latency and total usage rather than assuming more agents are faster or cheaper.
Implementation contract
route: orders | billing | technical | clarify allowed order tool: read_order_status financial actions: human approval required retry limit: 2 retries, 3 total attempts merge rule: all required checks must pass stop: validated output or explicit escalation
Record the route, not just the answer
Log the selected path, tool name, validated parameters, attempts, validation outcome and stopping reason. Avoid logging personal data or credentials unnecessarily. Correlate the steps with a run identifier so a reviewer can reconstruct the decision.
Explore the trade-off
Run the same fictional request through the router and sequential options in the lab. The fixture response may be identical while the trace shows different coordination.
Related: Test failure paths · Architecture library