# How Perceivo Orchestrates Why Perceivo builds a workflow from swappable components, and why that matters once a step uses a language model. - Page: https://perceivo.ai/university/orch-01-how-perceivo-orchestrates - JSON: https://perceivo.ai/registry/film/orch-01-how-perceivo-orchestrates.json - Kind: film - Revision: 3 (published 2026-10-18, produced 2026-10-15) - Previous revision 1 (published 2026-10-07): https://perceivo.ai/registry/film/orch-01-how-perceivo-orchestrates/r1.json - Previous revision 2 (published 2026-10-11): https://perceivo.ai/registry/film/orch-01-how-perceivo-orchestrates/r2.json - SDK version: none ## Validation - SDK: not applicable (Concept film; no component or orchestration is offered for download) - Platform: not applicable (Concept film; no component or orchestration is offered for download) ## Transcript [0:00] Introduction to Orchestrations, Episode 1: How Perceivo Orchestrates [0:06] Welcome to our Introduction to Orchestrations series. [0:09] Let's start with a simple idea. [0:11] An orchestration is a series of steps, executed in order. [0:16] Here's one you already know: an online store's checkout. [0:20] The cart is totaled. [0:22] The payment is charged. [0:24] The order is confirmed. [0:25] And a receipt goes out. [0:28] Every step is deterministic, meaning that every order is processed the same way, every time. [0:35] Even if you add a path for a failed payment, you are still going to end up on that path with one hundred percent certainty, if the payment declines. [0:43] Perceivo can run sequential flows like this one, step by step. [0:48] But the reason to build it as an orchestration, to separate it into building blocks, isn't just to illustrate its composition. [0:56] It's to isolate the individual operations. [1:00] Each step is its own component, with its own inputs and outputs. [1:04] And that means, you can take one out, and put a different one in. [1:09] For instance, we could swap the payment processor. [1:13] Run both versions side by side. [1:15] Compare them. [1:16] Nothing else in the flow changes, and that allows us to examine the impact of replacing a single component. [1:25] This kind of visibility, and comparability, become critical with orchestrations that have components that are not deterministic. [1:34] Say the store adds a chatbot to answer product questions. [1:37] It's powered by a language model, and language models are probabilistic. [1:43] So if you ask it the same question three times, it is possible you will get three different answers. [1:49] Two might be fine. [1:50] But the third invents a return policy the store doesn't have. [1:55] So we add a gate. [1:57] A second component checks each answer against the real policy. [2:01] If it passes, it goes to the customer. [2:04] If it fails, it goes back for another try. [2:08] That loop is iteration. [2:10] And it's where orchestration starts to earn its keep. [2:14] Iteration in business is rarely a straight line. [2:17] Take the store's ad script for a new product. [2:21] Sales writes a draft. [2:23] Legal rejects it. [2:24] Sales revises. [2:25] Legal approves. [2:26] Then branding has its say. [2:28] And then, maybe you get a result you can evaluate. [2:33] But would it go better if legal went first? [2:36] Or branding? [2:37] Same people. [2:38] Same components. [2:39] A different order. [2:40] You can't know which one is best until you try them, and measure. [2:46] And the answer doesn't stay put. [2:48] Regulations change. [2:49] Markets change. [2:51] The order that won last quarter may not win this one. [2:55] So you keep testing. [2:57] That's why Perceivo uses finite state machines to control how work moves. [3:03] A finite state machine is defined by a set of states, and the rules governing the transitions between them, known as a policy. [3:12] A state is where the work is right now, and what it's carrying, and the policy determines what state to move to next. [3:19] Perceivo treats policies as separate components that can be isolated, benchmarked, and replaced. [3:27] A lot of familiar designs are just different state machines. [3:31] A conditional router is a finite state machine that implements an if-then policy, sending work down one path or another. [3:39] A generator and discriminator: one component produces, another judges, and the loop runs until the work passes. [3:48] A mixture of experts: several specialists, and a policy that decides whose answer to trust. [3:54] And hierarchical reasoning: state machines inside state machines. [3:59] Higher layers make the big decisions. [4:02] Lower layers do the detailed work. [4:06] And there are many more. [4:08] A cascade, that tries a small model first and escalates when it has to. [4:14] A debate. [4:15] A planner and an executor. [4:17] An ensemble vote. [4:19] A human in the loop. [4:21] These are the ones we know. [4:23] The interesting ones are the ones nobody has drawn yet. [4:27] Let's build one, a piece at a time. [4:30] A personal shopping assistant. [4:32] It starts as a chatbot. [4:34] A customer asks, and it answers. [4:37] Sometimes the answers are wrong. [4:39] So we add a discriminator to check them before they go out. [4:43] Now it's accurate, but it isn't selling. [4:46] It doesn't know the customer. [4:48] So we connect their purchase history. [4:51] That tells us what they bought. [4:52] Not what they want. [4:54] So we add a recommender: a trained model that predicts what each customer is likely to want next. [5:00] But some customers aren't shopping at all. [5:03] They need to return a product. [5:05] So we put another trained model, a classifier, in front of our sales path. [5:10] It sends shoppers to the shopping assistant. [5:13] But everyone else goes to customer support, which in this case, we have incorporated as a nested orchestration inside this one node. [5:22] Every piece we added was a component we could isolate, measure, compare, and swap. [5:28] That allows us to evaluate changes in our architecture across a single dimension. [5:34] That's how Perceivo orchestrates. [5:37] And that is how orchestration helps you improve outcomes. [5:42] Everything in this series is open source, with every example in the Perceivo SDK repository. [5:49] Next time, we open the builder and take a tour of orchestrations. [5:54] See you there.