A buyer deciding how to orchestrate AI agents on Devotel Orbit is really choosing between four patterns: one canonical agent with a well-scoped prompt, a squad that routes each turn through a classifier to a specialist, a champion-challenger rollout that promotes prompt changes behind a quality gate, and role-play rehearsal that keeps human agents match-fit alongside the AI ones. Every one of those patterns has a shipped surface under /agents in the dashboard, and the posts in this cluster each cover one surface in depth. What has been missing is the index: which pattern, for which problem, decided how. This post is that index.
Updated 2026-09-17 with the definition section (what orchestration is, and the boundary it manages between deterministic flows and LLM agents), four full pattern subsections deep-linking the published walkthroughs, and an at-a-glance decision table.
What orchestration is: routing between deterministic flow logic and LLM agents
Orchestration has one decisive question at every step of a conversation: does this step need a language model's open-ended judgment, or is a deterministic flow node the right executor? A flow is code-backed and repeatable — it does the same thing on every run, which is exactly what you want for a balance lookup, a quiet-hours check, or a compliance gate. An LLM agent is probabilistic — it reads the customer's actual words, picks a model per policy, invokes tools, and composes an answer. Orchestration is the routing discipline that moves a conversation between those two execution models, preserving state across the handoff: flow → agent → escalation rules → human queue. Not every conversation needs an agent — a well-declared flow covers the predictable volume; not every conversation should stay in a flow — real customer language drifts off hardcoded branches. The what is LLM orchestration explainer walks the seven-step pipeline that routing runs (intent classification → model selection → tool dispatch → guardrails → synthesis → delegation); the four patterns below are the shipped shapes that pipeline takes.
The four shipped patterns, each walked end to end
Every pattern below is live on Devotel Orbit today — this index caps at shipped shapes, no roadmap.
Pattern 1 — Voice-agent ↔ flow handoff (when the conversation crosses execution models)
The workload: a real-time voice call where part of the run is deterministic (verify the caller, play the compliance notice, route by business hours) and part is open-ended (the customer's actual question, answered against a knowledge base with citations, with a delegation escape when a human should take it). The runtime executes the orchestration pipeline on the call — intent classification, model selection, tool dispatch, guardrail enforcement — and hands the run to a sibling agent or a human mid-call with a context note attached, rather than annotating the handoff after the fact. Start here when the conversation crosses between flow logic and an agent, either way. Deep dive: what is LLM orchestration.
Pattern 2 — Squads: multi-agent routing behind one conversation
The workload: inbound volume that has visibly split into distinct intents — billing, support, sales — where one prompt carrying every role's instructions starts to bloat and the guardrails muddy each other. A classifier agent reads every inbound turn and routes to the specialist member that owns that intent, with an optional fallback for unmatched labels, all behind one conversation record and with per-member routing analytics. Pick it when the intent set competes for the model's attention on every turn, or when different roles need different model tiers. Deep dive: agent squads — multi-agent orchestration.
Pattern 3 — Canary rollouts: champion-challenger promotion behind a quality gate
The workload: a prompt that is already live and earning, whose next revision is non-trivial. Instead of swapping at 100% and learning about the regression from customers, the candidate version takes a percentage ladder — 5% → 25% → 50% → 100% by default — advancing only while a six-metric quality scorecard holds, and pulling back automatically on any regression. Use it for every non-trivial change to a live prompt; the cost of caution is a few days, the cost of a bad swap is trust. Deep dive: canary rollouts — champion-challenger promotion.
Pattern 4 — Flow routing with intent classification (LLM judgment inside a deterministic graph)
The workload: routing decisions where customer language drifts off the keywords a code-based branch hardcodes. The flow builder's AI Classify node declares the intents you serve — "billing", "support", "booking" — picks the one a message actually means, and routes down a named edge per intent, with a designed fallback path when nothing matches. The routing stays observable: the flow's step trace records which intent matched, which edge the run took, and when the fallback fired. Use it when the routing decision must live inside a deterministic flow but keyword matching keeps failing on real language. Deep dive: intent classification vs code-based routing.
The decision table: which pattern fits which workload
| Your workload looks like | The pattern to pick | Deep-link walkthrough |
|---|---|---|
| A voice call that crosses between deterministic steps (verification, compliance gates) and open-ended answers, with a human handoff option | Voice-agent ↔ flow handoff | What is LLM orchestration |
| Inbound volume split into distinct intents (billing / support / sales) that one prompt can no longer hold | Squads — classifier + specialists | Agent squads |
| A live prompt whose next revision is anything beyond a typo fix | Champion-challenger canary rollout | Canary rollouts |
| A routing decision inside a flow where keyword branches keep missing real customer language | Flow routing with intent classification | Intent classification routing |
| A team whose QA finding is a skill gap, not a defect | Role-play rehearsal in the Studio (layered on any pattern above) | The Studio |
Grounding and payment mandates cut across every row of the table — they constrain and de-risk whichever pattern you pick, they do not decide its shape.
The shipped surfaces and the pattern each one carries
The orchestration palette lives under /agents. Four surfaces matter most when you are deciding on a pattern rather than configuring one:
- [The Marketplace](/agents?tab=marketplace) carries the start-from-a-proven-shape pattern. Published agents arrive with tools, rubrics, and a reference prompt already wired, so the first deployment decision is "which shipped shape is closest to my workload" rather than "what do I write in a blank prompt field."
- [From Prompt](/agents/from-prompt) carries the single-canonical-agent pattern. Describe the agent in plain English; the builder drafts the name, system prompt, the smallest tool set from the registry, evaluation rubrics, and seeded regression inputs, and the draft is sandbox-testable before it ever touches a customer. One agent, one prompt, one knowledge base — the right default until it isn't.
- [The Studio](/agents?tab=studio) carries the rehearse-before-production pattern. Supervisors author scenarios and human or AI agents rehearse against an AI customer with an automatic score on every run, so a weak handoff or a thin guardrail is found in practice rather than in a live queue.
- [Squads](/agents/squads) carries the classifier + specialists pattern. A classifier agent reads every inbound turn and routes to whichever specialist member owns that intent, with an optional fallback agent for unmatched labels — one conversation record, multiple agents behind it.
Three adjacent patterns complete the picture and have their own surfaces: [canary rollouts](/blog/agent-canary-rollouts-champion-challenger-2026) (champion-challenger promotion of a candidate prompt version through a percentage ladder with a quality scorecard gating each step), [RAG grounding](/blog/rag-grounding-citations-ai-agents-2026) (knowledge-base scope plus a citation directive so answers carry their sources), and [payment mandate guardrails](/blog/agent-payment-mandates-vs-category-2026) (decision-level gates on what an agent may commit to financially). These are depth patterns you layer on top of whichever structure you pick — they constrain and de-risk the agent; they do not decide its shape.
When the pattern beats the alternative
The decision is rarely "AI or not AI" — it is which orchestration shape carries the workload safely. The four-way decision rule:
- One canonical agent beats a squad when your intent set is small and stable. If one prompt can hold the scope limits, the escalation rules, and the tone for everything the agent touches, a squad adds routing machinery that solves a problem you do not have. Start here; the from-prompt builder makes the start cheap.
- A squad beats one canonical agent the moment the single prompt starts to bloat — when billing, support, and sales instructions compete for the model's attention on every turn, when one role needs a heavyweight model and another should run a cheap one, or when "issue a refund" and "browse the catalog" share a prompt and the guardrails muddy each other. Specialization resolves all three; the squads post walks the mechanics.
- A canary rollout beats swap-at-once whenever the prompt is already live and the change is non-trivial. Swapping a production prompt at 100% means you learn about the regression from your customers. The champion-challenger ladder (5% → 25% → 50% → 100% by default) advances only while the six-metric scorecard holds, and pulls the candidate back on any regression — the cost of caution is a few days, the cost of a bad swap is trust.
- Rehearsal beats QA-on-live-traffic when the finding is a skill gap rather than a defect. Scoring a live call tells you who needs training; it does not make them better. The Studio's role-play turns the QA finding into a practice loop, so the next live conversation is not the rehearsal.
Grounding and payment mandates cut across all four: an ungrounded specialist inside a well-designed squad is still an ungrounded answer, and a squad member that can issue refunds needs the mandate gate regardless of how clean the routing is.
Worked scenarios, one per pattern
Route a support queue with a classifier, escalate to specialists. A tenant running general inbound support promotes three drafts out of from-prompt — a "Front door" classifier, a Support specialist, a Billing specialist — and composes them in Squads with a fallback catch-all agent. A customer writes "My card was charged twice"; the classifier returns the billing intent and the Billing specialist, grounded on the billing knowledge base, resolves it in the same conversation record. Later turns about upgrading route to the Support specialist without the customer noticing a handoff — and the squad analytics show whether the fallback is doing too much work, the signal that the intent labels need a pass.
Canary-roll a new prompt to 5% before the swap. The Support specialist's prompt gets a rewrite for seasonal volume. Instead of swapping, the operator opens a canary rollout on the candidate version: 5% of traffic sees the challenger, the six-metric scorecard compares it against the champion, and the ladder advances only while the candidate holds. A tone regression at 25% pulls the candidate back automatically — the fix happens in the rollout surface, not in a customer-facing apology. The full ladder and the guardrail-gated promotion mechanics are in the canary rollouts post.
Rehearse the human side before the queue does. QA on live calls flags a team's thin de-escalation handling. The supervisor authors a de-escalation scenario in the Studio, agents rehearse against an AI customer that does not calm down on the first attempt, and every run comes back with an automatic score. The next live escalation starts with the representative already having lost that argument three times in practice instead of zero.
Ground the answers before publication. Before any specialist in any of the above goes live, its answers go through the grounding loop — knowledge base scoped, citation directive in the guardrails, links surfaced per turn, and the grounding-rate audit watched on the dashboard — the pattern the RAG grounding post walks end to end.
Frequently asked questions
Which pattern should a team deploying its first agent start with?
One canonical agent, drafted through the from-prompt builder. It produces the prompt, tools, rubrics, and seeded test inputs in one pass, and the draft is sandbox-testable before activation. Graduate to a squad when the single prompt visibly bloats or the routing analytics say the workload has split into distinct intents.
When does a squad beat a single well-prompted agent?
When one prompt carries instructions for multiple functions and the symptoms appear: prompt bloat degrading answers, one model forced to serve both cheap FAQs and heavyweight reasoning, guardrails from different roles interfering, and no per-role analytics. A classifier plus specialists resolves each of those while keeping one conversation record.
Why not swap a new prompt version at 100% and watch the dashboard?
Because watching the dashboard is how you find out from your customers that the swap regressed. The champion-challenger canary ladder starts at 5% and advances only while the quality scorecard holds, pulling the candidate back on any regression — the same information, without the blast radius.
What does the Studio rehearse — AI agents or human ones?
Both. Supervisors author scenarios, and either a human agent practicing de-escalation or an AI candidate being exercised before promotion runs against an AI customer with an automatic score on every run. The pattern is the same: find the weakness in rehearsal, not in the queue.
Do grounding and payment mandates replace the structure patterns?
No — they layer on top. Grounding constrains what an agent may claim; payment mandates constrain what it may commit to. Both apply whether the agent is a singleton, a squad member, or a canary candidate, and both are configured per agent in its guardrail settings.
Where do I see whether the pattern I chose is working?
Per surface: squad routing quality on the squad's analytics tabs (per-member utilization, handoff success, drop-off), canary candidates on the rollout scorecard, grounding on the grounding-rate audit, and rehearsal on the Studio run scores. Each pattern has its own feedback loop; none of them require inferring health from raw volume.
The takeaway
Orchestration on Orbit is not one feature — it is a palette of shipped patterns with a decision rule attached: one canonical agent until the prompt bloats, a squad when routing becomes the problem, a canary ladder whenever a live prompt changes, rehearsal whenever the gap is skill rather than configuration, and grounding plus mandates across all of them. The surfaces under /agents — Marketplace, From Prompt, Studio, Squads — are where each pattern starts, and the posts in this cluster are where each one is covered in depth.