Tutorials

Tutorials walk you through complete, end-to-end workflows using SOAT. Each one starts from scratch and demonstrates how the platform's building blocks fit together in a real scenario — every step comes with CLI, SDK, and curl examples, and every tutorial is validated against a live server in CI.

Before you begin

All tutorials assume:

A running SOAT instance. Follow the Quick Start to bring the stack up.
Familiarity with SOAT's core concepts. Read Key Concepts if you are new.
The CLI installed, or the SDK set up in a TypeScript project.

Pick a path

New to SOAT? Start with Permissions in Practice and Chat with an LLM, then connect a hosted model with Connect Third-Party LLMs.
Building an agent? Wire platform tools in Agent SOAT Tools, run your own functions with client tools, add persistent memory, and learn to debug sessions, generations, and traces.
Locking an agent down? Cap what a single agent may do — independently of who calls it — with Bound an Agent with a Boundary Policy, then gate the tool types a boundary cannot reach using Guardrails.
Retrieval and RAG? Generate embeddings, build an agent over a library of PDFs, and ingest images and audio.
Coordinating multiple agents? Compare a direct orchestration pipeline, a workflow state machine, and nested agent calls — then add branching, control flow, or human approval gates.
Building one governed process end to end? Close the Monthly Books composes an orchestration, a workflow, a trigger, and a human approval into a single month-end close — the capstone that puts the pieces above together.
Shipping to production? Deploy a whole stack declaratively with Formations and an agent squad, automate it with Triggers, gate risky calls with Guardrails, hand per-user credentials to tools with tool context, and control cost with metering and budgets and per-end-user spend caps.
Shipping a change safely? Measure it first with Evaluate an Agent, grade open-ended output with an LLM judge, then make the rollout itself wait for a green suite with Gate a Canary Promotion on an Eval.
Chasing one bad answer? Replay a Bad Turn reads the turn back step by step, freezes it as a fixture so it cannot regress quietly, and forks the session at that message to try a different agent against the same context.
Under compliance pressure? Erase, expire, or never store prompt content with data retention and zero-retention, and ship prompt changes safely with agent versioning and canary rollout.
Calling a cloud API? Sign requests as AWS or authenticate as a GCP service account with Call AWS and GCP APIs from an Agent.