TECHNIQUE
Agentic Orchestration
Observed routing / intent classification is usually an upstream control point that sends a request or code change into a bounded domain, backend, tool, or specialist agent before generation/retrieval/review.
Put an upstream router/classifier in front of the main workflow: Dropbox has a classifier pick a sub-agent; Halliburton’s Intent Router classifies an intent label; LinkedIn classifies user intent and routes queries; Rippling’s supervisor decides which specialized agent or combination to invoke; Uber’s intent agent maps questions to business workspaces; DoorDash routes PR review context to domain profiles.
6 of 12 roster operators have cited routing evidence in this pool; among evidence-backed operators, 6 of 6 show an upstream router/classifier.Use routing to narrow search radius or context before downstream work. Uber maps the question to business domains/workspaces to narrow RAG; DoorDash loads only the relevant domain review rules; Rippling identifies the relevant domain and injects a scoped skill; LinkedIn constructs retrieval strategy and routes to the most efficient graph or GraphQL backend.
4 of 12 roster operators have cited evidence for routing as context/search narrowing; 4 of 6 evidence-backed operators show this pattern.Route into specialized agents, tools, or handlers rather than relying on one monolithic prompt. Dropbox uses sub-agents for complex queries; Rippling coordinates 5 to 7 specialized subagents; Halliburton separates Q&A and workflow generation with a tool-bound agent; LinkedIn uses tool/function calling to choose graph node types or retrieval actions; Uber splits QueryGPT into intent, table, and column-pruning agents.
5 of 12 roster operators have cited evidence for routing into specialized handlers; 5 of 6 evidence-backed operators show this pattern.Evaluate routing pipelines with regression suites, production scoring, or component-level tests. Dropbox tests a chain that includes intent classification; Uber evaluates QueryGPT components and tracks regressions; Halliburton built SME-validated query-workflow tests; DoorDash uses evals plus production acceptance; Rippling uses traces, layered evals, and production monitoring; LinkedIn’s SPP AI includes an accuracy testing framework.
6 of 12 roster operators have cited evidence for evaluating systems that include routing/classification; 6 of 6 evidence-backed operators show evaluation or monitoring around the routed workflow.Every evidence-backed operator places routing/classification inside a larger multi-stage workflow, not as a standalone product feature: examples include PR review, enterprise search, seismic workflow generation, job search, security query answering, workforce-management agents, and text-to-SQL.
Every evidence-backed operator routes into a bounded inventory of destinations or scopes: domain profiles, sub-agents, intent labels, graph node types/backends, workspaces, or specialized agents.
The routed destination differs by deployment.
APPROACH 01
Route to domain or context scope before continuing.
APPROACH 02
Route to specialized agents or sub-agents.
APPROACH 03
Route to data backends, graph node types, or retrieval strategy.
The decision mechanism differs.
APPROACH 01
Explicit LLM or classifier produces an intent/category/routing decision.
APPROACH 02
Supervisor-agent reasoning loop decides which specialized agent or combination to invoke.
APPROACH 03
Route from the touched code/domain to prewritten review profiles and load only those rules.
Routing mistakes can cascade because downstream generation, retrieval, ranking, prompting, and safety/filtering depend on earlier stages; operators call out unpredictable pipeline ripple effects, attention/focus problems, and domain disambiguation across large schemas or ontologies.
Evaluation and feedback loops are needed to keep routed systems useful: operators report regression tests, production scoring, SME validation, incident evals, component-level evaluation, and suppression of historically low-value categories.
Latency, cost, and compute pressure shape routing design and model choice: Halliburton selected Nova Lite for routing accuracy/latency balance, DoorDash stages cheaper and stronger models, and LinkedIn reports advanced LLM compute requirements plus serving optimizations.
The destination inventory behind the router requires curation and upkeep: Uber uses curated workspaces and manual golden question-to-SQL mappings, DoorDash maintains per-domain review profiles, and LinkedIn uses human annotations to train policy behavior.
| Name | Kind | When | Maturity |
|---|---|---|---|
| Constrained JSON-schema classification on a fast model | pattern | routing decisions need typed enum outputs at low latency and cost | commodity |
| Embedding-similarity routing | pattern | route by nearest exemplar without an LLM call per request | established |