Examples and recipes

Start from a working program with a visible result. Join at the step you need: the projects below are independent, and you do not have to complete them in order.

Explore a harness locally

Approve or deny a simulated refund. Download, install, and run without credentials.

Build a repeatable quality check

Run passing and failing cases, inspect traces, then opt into a live model.

Choose a complete project

Every downloadable project includes TypeScript source, dependencies, a compiler configuration, and check and start scripts. Read its execution mode before running it.

Build toward	Project and result	Packages	Runs with
First response	Create an Agent: scaffold, run, change the prompt	Core + CLI scaffolding	OpenAI credentials
Application capabilities	Support checkpoints: inspect tools, sessions, evidence, approvals, and HTTP independently	Core; transport for HTTP	Model credentials; embedding access for retrieval
Controlled research	Research harness: a tool result becomes a cited answer	Core + harness	Model credentials; local policy data
Conversation continuity	Harness sessions: two turns, four history messages, a separate user’s empty history	Core + harness	Model credentials
File-based context	Workspace reviewer: find an unfinished task in selected files	Core + harness	Model credentials; temporary local files
Approval before an effect	Refund workflow: denial makes zero submissions; approval makes one simulated receipt	Core + harness	Local, no credentials
Retained graph output	Execution graph: dependency calculation, completed-task reuse and explicit root limits	Core + harness	Local, no credentials; process-local store
Repeatable quality	Quality gate: passing and failing cases, exit codes, report and traces	Core + eval + observability	Local fixture; explicit optional live mode
Authenticated delivery	Owned-session API: verify 401, 200, and cross-user 403 responses	Core + transport	Loopback HTTP fixture; explicit optional live mode
Background work	Producer and worker: enqueue, inspect the terminal result, drain	Core + queue	Local Redis; deterministic model
Phone call control	Call intent lifecycle: deduplicate, authorize, and cancel a call intent	Core telephony entry point	Local carrier fixture, no actual call

The call-intent project demonstrates call control. For a spoken telephone assistant, also connect a media path and speech adapter. The support chapters are independent learning checkpoints, not a single cumulative application.

Adapt a focused pattern

These pages assume an existing application. Their functions identify the model, data, services, or runtime the host supplies; they are not all standalone projects. Use Build when you need help choosing the next capability.

Task	Patterns
Configure an Agent	Agents, lifecycle, skills
Keep and scope context	Memory and sessions, storage, learning
Use tools and evidence	Tools, toolkit directory, retrieval, MCP
Select models	Selection and fallback, provider setup, Jev
Coordinate work	Teams, workflows, handoff
Run work later	Queues, scheduling
Connect applications	Transport, integrations, webhooks
Work with media and devices	Voice and browser, edge
Measure behavior	Evaluation, cost and observability
Apply controls	Security, production patterns, advanced capabilities
Study a larger composition	End-to-end patterns

What is verified

The eight downloadable projects target the 4.9 release candidate. Local fixture checks exercise the harness projects, graph dependencies and retained output, both quality-gate outcomes, HTTP identity and ownership, and simulated telephony. The queue project has a separate disposable-Redis smoke test. The graph fixture is process-local and does not establish crash durability. These checks do not establish registry publication, live provider or carrier behavior. Displayed source and project downloads are generated from the same files. Model-based projects accept OPENAI_MODEL; see example model choices for documented defaults. The separate examples repository contains more applications; check their package versions before adopting one.

Complete projects

Agents & memory

Models & Jev

Tools & knowledge

Teams & jobs

Deploy & devices

Quality & production

Examples and recipes

Explore a harness locally

Build a repeatable quality check

Choose a complete project

Adapt a focused pattern

What is verified

Explore a harness locally

Build a repeatable quality check

​Choose a complete project

​Adapt a focused pattern

​What is verified

Choose a complete project

Adapt a focused pattern

What is verified