examples/

Command agent-approval gates a dangerous tool behind human approval: the gate pauses the run, the operator answers on the terminal, and a second Run resumes the conversation with the outcome.

Command agent-basic runs an agent loop end to end: the model calls a typed tool, the runtime executes it and feeds the result back, and the loop ends when the model answers in plain text.

Command agent-harness demonstrates the stateful harness layer: a session persisted as a JSONL file survives process restarts, prompts reconstruct their context from the tree, and oversized history is compacted automatically.

Command agent-stream consumes an agent run as a live event stream: model text deltas print as they arrive, and tool lifecycle events interleave in real time.

Command agent-subagent nests one agent inside another as a tool via agent.AsTool: the outer agent delegates research questions to an inner agent bound to its own prompt, each invocation running in an isolated session.

Command agnes-image generates and edits images with the Agnes Image API.

Command image-gen generates an image with gpt-image-2 and writes it to disk.

Command middleware shows composing a bare provider with rate limiting, retries, and observability via ai.Chain.

Command provider-switch sends one prompt through any built-in provider package while keeping the portable ai.Request unchanged.

Command rerank demonstrates document reranking with the Cohere adapter and compatible services (SiliconFlow, Jina AI, Together AI, or local vLLM/TEI).

Command structured extracts typed data from free text with ai.GenerateTyped; the JSON schema is derived from the Go struct.

Command team-agent 演示如何使用独立 Agent Session、Continuation 和 Team 组合出一个可恢复的多 Agent 协作流程。

Command text-stream-anthropic streams a message from Anthropic Claude.

Command text-stream-gemini streams a response from Google Gemini.

Command text-stream-openai streams a chat completion from OpenAI and prints tokens as they arrive.

Command tools runs a complete tool-calling round trip: the model requests a tool invocation, the program executes it and returns the result, and the model produces the final answer.

Command vision sends an image to a vision-capable model and asks for a description.