Runnable Examples

The guides contain complete inline examples, including setup, invocation and observable results. The downloadable scripts below are optional companion files. Start with Your First Agent and Build a Support Agent if you are new to the library.

Find an inline example

Run an example

From a repository checkout, install dependencies and configure a provider key in your environment or the checkout's untracked .env. The example runner loads that file; a consuming application configures credentials itself.

mix deps.get
mix examples.live --model openai:gpt-5.4-mini --only examples/react_agent.exs
mix examples.live --provider google --only examples/streaming_agent.exs

The weather example calls a local fixture tool with a real model. It proves the model/tool integration, not access to a weather service. Script comments identify fixture-backed business tools and required external resources.

Choose models explicitly with --model; use --provider for runner defaults. TypeSafe runs decision-specific examples rather than generic chat scripts. Some scripts require more than one provider, as stated in their setup.

Run a selection or the full matrix

mix examples.live --provider openai,google
mix examples.live --provider openai --offline
mix examples.live --provider moonshot --only deepagents/ --timeout 600
mix examples.live

The unfiltered command requests every manifest provider and fails for missing credentials. --jobs controls concurrency and --timeout is seconds per script. Generic live scripts run across applicable chat providers; pinned scripts use their actual provider. Offline scripts exercise local behavior. Transient provider failures can receive one bounded retry; failed behavioral assertions are not silently converted into passes.

Understand verification results

Each run retains per-attempt logs, JSON results and provider HTTP evidence under _build/examples_live/runs/. A successful process exit is insufficient: a live example must receive a successful inference response from its required provider and pass its behavioral assertions. Unconsumed streams, constructors, model lists and fixtures cannot substitute for inference.

Tool assertions check successful results and relevant arguments. Other checks cover structured answers, streamed events, approval/resume, saved memory and actual files. A provider response alone does not establish all those behaviors.

The existing mix docs.verify --all command additionally executes guide and inline API examples and a fresh consumer of the built package. Use --only for a page/script substring, --scenario for a scenario, --provider or --model for provider/model selection and --service postgres,docker for service requirements. Missing prerequisites remain incomplete in reports. Public Hex installation is separate and remains pending until the target release is published and installed successfully.

Agents and tools

Memory, subagents and deep agents

The deepagents/ folder holds 15 more live deep agents: research with subagents, a coding agent with write permissions, text-to-SQL, a wiki keeper, a swarm of parallel workers, a recursive agent, async subagents and more. Its README lists each one with the tools it must call.

Graphs

Retrieval, streaming and sandboxes

Provider features

Choose a useful integration check

Use the actual provider/model, storage adapter and execution backend that you deploy. For media examples, use the bundled fixtures and create/delete remote resources as the script demonstrates. For retrieval, assert expected sources with real embeddings. For caching, inspect observed cache reads; zero is not a cache hit. For generated work, inspect the actual artifact.

Testing Your Application covers deterministic application tests; Production Integration covers deployment ownership.