Loading...
Loading...
Found 13 Skills
Pause an AG2 `Agent` mid-run to collect human input via `context.input()`, or gate a tool call with `approval_required()` middleware. Use when the user wants the agent to ask for confirmation, request missing info (passwords, API keys, data), or have a human approve sensitive / irreversible / expensive tool calls (sending emails, deleting records, payments).
Send images, audio, video, or documents into an AG2 `Agent` alongside text. Pass `ImageInput`, `AudioInput`, `VideoInput`, or `DocumentInput` as positional args to `agent.ask(...)`. Use when the user wants the agent to process non-text input — describe a photo, transcribe audio, summarise a PDF, analyse a video. Covers per-provider support matrix, the four ways to source data (URL / path / bytes / file_id), Gemini-specific YouTube + media-resolution + clipping, OpenAI image-detail, Anthropic prompt-caching on attachments, and `FilesAPI` for upload lifecycle.
Host an MCP server that exposes an AG2 `Agent` (plus prompts and resources) to MCP clients like Claude Desktop, Cursor, or the MCP Inspector. Wrap the agent with `MCPServer(agent)` — it surfaces `Agent.ask()` as a single conversational tool and serves over stdio (`run_stdio()`) or streamable HTTP (it is itself an ASGI app for uvicorn). Covers `MCPServer`, `SessionConfig` (multi-turn history), `Prompt`/`PromptArgument`/`PromptMessage`, `Resource`/`ResourceTemplate`, `AskContext`/`ContextProvider` (per-request injection), `build_ask_tool`, OAuth2 `security=`, and in-process `testing.connect`/`testing.serve` helpers. Use when you want OTHER MCP clients to call YOUR agent. This is the SERVER side — for CONSUMING external MCP servers from an agent (client side) see `ag2-use-builtin-tools` (`MCPServerTool`).
Map of AG2 capabilities and which sibling skill to reach for. Load first when the user mentions building with AG2 (ag2) but the specific feature isn't yet clear — agents, tools, model config, delegation, memory, observers, structured output, HITL, AG-UI, MCP server hosting, A2A protocol, realtime voice, telemetry, testing, or evaluation.
Drive external CLI coding agents (Claude Code, Codex, OpenCode) as first-class AG2 Agents over the Agent Client Protocol (ACP): point an Agent at a ClaudeCodeConfig / CodexConfig / OpenCodeConfig preset from ag2.acp and ask()/run() it like any agent, with its thinking, tool calls, plans and permission prompts externalized onto AG2's event stream. Use when orchestrating, observing, or gating CLI coding agents from Python — permission_policy ask/auto/deny HITL gating, fs_root confinement, turn timeouts, in-process testing via fake_acp_config. To expose an AG2 agent to other systems instead, see ag2-a2a or ag2-mcp; for approval plumbing see ag2-hitl.
Single-agent recursion and parallel fan-out within one AG2 `Agent` — auto-injected `run_subtask` / `run_subtasks(parallel=True)` (opt in via `tasks=TaskConfig(...)`) for self-delegation, and `Agent.as_tool()` as a lightweight no-hub way to call one named agent from inside another. Use when a single coordinator wants to break work into its own sub-tasks, fan out concurrent sub-tasks, or invoke a specialist agent as a tool. Covers context flow, recursion safety, and `persistent_stream` for sub-task history. **For two or more agents actually collaborating with a registry, durable channels, governance, or turn-taking, use `ag2-network-quickstart` instead** — the network is the standard multi-agent pattern in AG2.
Wire AG2's shipped tools into an `Agent` — both provider-native server-side tools (web search, web fetch, code execution, MCP, image generation, memory) and locally-executed common toolkits (filesystem, DuckDuckGo, Exa, Tavily, skills). Use when the user wants capabilities AG2 already ships rather than writing custom Python. For shell commands see `ag2-shell-tool`; for custom Python tools see `ag2-add-custom-tool`.
Build a minimal AG2 `Agent` end to end — pick a model provider, set a prompt, call `agent.ask()`, then continue the conversation with `reply.ask()` (multi-turn). Use when the user is starting a new AG2 project, has no working `Agent` yet, or needs the multi-turn chaining pattern. Covers `OpenAIConfig`, `AnthropicConfig`, `GeminiConfig`, `OllamaConfig` etc., and env-var fallback for API keys.
Expose an AG2 `Agent` over the AG-UI protocol so a frontend (CopilotKit, custom React/Next.js, or any AG-UI client) can stream responses (text and reasoning), render tool calls, sync shared state, and track sub-task steps. Wraps the agent with `AGUIStream(agent)` and mounts it in FastAPI via `stream.dispatch(...)` or `stream.build_asgi()`. Use when the user wants a web frontend in front of an AG2 agent rather than a CLI / script.
Monitor an AG2 agent's stream — log events, detect repeated tool calls, track token spend, build trigger-driven observers, route observer alerts to the model, and halt on FATAL conditions. Covers `@observer(...)` (stateless), `BaseObserver` (stateful), built-ins (`TokenMonitor`, `LoopDetector`), `Watch` primitives (`EventWatch`, `CadenceWatch`, `DelayWatch`, `IntervalWatch`, `CronWatch`, `AllOf`, `AnyOf`, `Sequence`), `ObserverAlert` (`Severity.INFO/WARNING/CRITICAL/FATAL`), `AlertPolicy`, and `HaltEvent`. Use when the user wants observability, runtime safety guards, alerts, or batch/time-based reactive logic.
Add OpenTelemetry traces to an AG2 `Agent` via `TelemetryMiddleware` (`ag2.middleware.builtin`). Emits spans for the full turn, each LLM call, each tool execution, and each human-input request, following the OpenTelemetry GenAI semantic conventions. Compatible with any OTLP backend — Jaeger, Grafana Tempo, Datadog, Honeycomb, Langfuse. Use when the user wants production-grade traces, latency analysis, token-usage attribution, or to ship telemetry into an existing observability stack.
Get a typed Python value back from an AG2 `Agent` instead of free text. Pass `response_schema=` (a Pydantic model, dataclass, primitive, union, `ResponseSchema`, or `@response_schema` validator) and read the parsed result via `await reply.content()`. Use when the user wants validated structured output, classification, extraction, or scoring. Covers `ResponseSchema`, `@response_schema`, `PromptedSchema` (for providers without native structured output), per-turn override, validation retries, and primitive embedding.