Recipes

Recipes are variations of the same harness: state is a record, one turn is a stream, and the loop decides what happens next.

Each recipe shows one design move you can copy into your own agent: gate tool calls before execution, stream inner work, catch provider failures, rewrite history, pause between turns, or fan out to multiple models. They are real, type-checked code in recipes/, and each page mirrors the README.md sitting next to its index.ts and tests.

For the foundational shapes, start with One turn is a stream, Basic usage, and Structured output.

Tools and HITL

Recipe	One-line
Tool call approval	Gate sensitive calls before Toolkit.run; still return one result per model-requested tool call.
Streaming tool output	Show inner tool work to the user while returning one clean output to the model.
Streaming structured output	Validate prompted JSONL one object at a time as the model streams.

Reliability and lifecycle

Recipe	One-line
Multi-model fallback	Recover from provider stream failures by advancing to the next tier.
Model retry	Add retry policy around one model stream; only transient provider failures get another try.
Auto-compaction	Rewrite oversized history as an ordinary state transition.
Pause and resume	Pause between loop iterations with a latch; no provider call remains open.
Mid-stream abort	Cancel an in-flight turn through stream interruption and scope cleanup.
Agentic loop	Drive a long-lived chat from a user-message queue while continuing model/tool work between clean turn boundaries.
Sleeper agent	Wait for a long-running tool call. The agent goes quiet while the work runs and wakes up when it’s done.

Observability

Recipe	One-line
Basic metrics	Stack time-to-first-token, throughput, token totals, and completion time onto a generation; log them live while the story streams to a file.

Transport

Recipe	One-line
Modify output stream	Keep the loop transport-agnostic; project typed turn events into SSE or JSONL at the edge.

Multi-model

Recipe	One-line
Model escalation	Start on a fast cheap model; let it escalate hard questions to a stronger tier via a tool call.
Multi-model compare	Fan one prompt out to multiple providers; per-member errors stay isolated.
Model council	Build a stream graph where models answer, judge each other, and emit a winner.

Web search

Recipe	One-line
Grounded answer	Answer a current-events question from live web search with inline citations; swap the LLM and backend at will.
Deep research	Plan a broad question into sub-questions, investigate each with a streaming sub-agent, synthesize a cited report.
Native deep research	Submit one question to a provider-hosted research agent (OpenAI, Perplexity, Gemini, Exa); stream its progress and get one cited report.

Web reading

Recipe	One-line
Market intel	Read a batch of vendor pages to clean markdown and extract a typed pricing record from each, concurrently; no selectors, so pages that share no layout still decode.

Speech

Recipe	One-line
Basic transcription	Transcribe a file via the generic Transcriber service; swap providers with --provider.
Basic speech synthesis	Synthesize a phrase via the generic SpeechSynthesizer service; sync or chunked-streaming mode.
Streaming transcription	Live mic → transcript over WebSocket; Bun server bridges browser AudioWorklet to provider realtime.
Streaming synthesis	Type text → audio plays as the first chunk arrives; incremental text-in over WS.
Voice loop	Full STT → LLM → TTS pipeline with stop-word interrupt and follow-up queueing; one fiber per turn.

Music

Recipe	One-line
Basic music generation	Generate a 30-second clip with Lyria 3; simple prompt or weighted prompts with lyrics and BPM hints.
Radio station	Run your own AI radio station. An AI DJ writes the next track while you listen to the current one; the same set replays for free after the first pass.

Sandboxes

Recipe	One-line
Run, fix, repeat	LLMs are bad at exact computation. Give them Python. Tracebacks from a sandboxed microVM feed back into the next turn.

Browser

Recipe	One-line
Agent usability testing	Give an agent a goal and a URL; it drives the site like a first-time visitor and reports where the UX tripped it up.
Dashboard briefing	Screenshot a dashboard whose charts only exist client-side and decode one vision turn into a typed briefing.