Agent
adk-go-workflows
Requires `google.golang.org/adk/v2 >= v2.0.0`, which is where the `workflow` package and `agent/workflowagent` first ship.
@googleView AgentDevelopment
Agent
live-eval
`--mode adk_live` runs an eval over the agent's `/run_live` WebSocket instead of `/run_sse`. Pass it on its own for a local autoboot, or alongside the agent's `https://` `--url` when deployed (for Agent Runtime, the full engine URL). The dataset, the trace output, and the `eval
Install
$ npx -y skills add google/agents-cli --agent claude-codeHow it fires
How this agent gets triggered: by you, by Claude, or both.
- Fires itselfAuto-invocation. Claude auto-loads it when your prompt matches the work.Auto-invocation is when the right skill fires by itself at the right moment, driven by a FLOW.md router and a hook, instead of you invoking it by name. It is the difference between a skill being installed and a skill actually getting used.Read the full definition →
- You can call itInvoke it directly when you want it.
Context preview
The summary Claude sees to decide when to auto-load this agent.
`--mode adk_live` runs an eval over the agent's `/run_live` WebSocket instead of `/run_sse`. Pass it on its own for a local autoboot, or alongside the agent's `https://` `--url` when deployed (for Agent Runtime, the full engine URL). The dataset, the trace output, and the `eval
Agent definition
live-eval.mdEvaluating Live agents
`--mode adk_live` runs an eval over the agent's `/run_live` WebSocket instead of `/run_sse`. Pass it on its own for a local autoboot, or alongside the agent's `https://` `--url` when deployed (for Agent Runtime, the full engine URL). The dataset, the trace output, and the `eval grade` step are unchanged.
# Local autoboot
agents-cli eval generate --mode adk_live
# Deployed agent
agents-cli eval generate --mode adk_live --url https://my-live-agent.run.app --app-name app
# Chain generate + grade
agents-cli eval run --mode adk_live --url https://my-live-agent.run.app --app-name app \
--metrics final_response_quality
What gets graded
Live replies are audio, and are transcribed by default, so the **transcript** is what gets graded. Raw audio and video bytes are dropped from the trace.
How cases are played
Each user turn is sent as **text** over **one persistent socket**, in order. A single-turn case (`prompt`) is just a one-turn conversation. History lives in the live session rather than being seeded over HTTP, so the model conditions on the real running conversation, and multi-turn cases produce a trajectory that `multi_turn_*` metrics can score.
Author **user-only** turns. The agent generates every reply over the live session, so pre-authored agent turns are ignored (the CLI warns) and are not seeded as history.
Two failures that look like transport bugs
Both let the socket connect and then fail inside the session, so the error points at `/run_live` when the cause is agent configuration.
| Cause | Fix | |---|---| | The agent is not on a Live model. `--mode adk_live` changes only the transport; it cannot make a non-Live agent bidi. The scaffold default is not a Live model, so a fresh project fails with WebSocket code 1011. | Switch the agent to a Live model. | | On Vertex, the model's region is not pinned. Live models are served from a regional endpoint such as `us-central1`, not `global`. Unpinned, the model falls back to `GOOGLE_CLOUD_LOCATION`, which `agents-cli deploy` sets to `global` on Agent Runtime. | Pin the region on the model itself rather than steering `GOOGLE_CLOUD_LOCATION`, which is shared with sessions, telemetry, and grading. |
Grading is unaffected by either: `eval grade` passes its own location (`--region`, default `global`), so `eval run --mode adk_live` chains both steps correctly in one command.
> **ADK projects.** Current Live model IDs, and the code to pin a model's > region, are in `/google-agents-cli-adk-code` (`references/adk-python-live.md`, > "Models"). Transport mechanics are on the same page under "Events" and > "Serving and testing a Live agent".
Read more
Evaluating Live agents
`--mode adk_live` runs an eval over the agent's `/run_live` WebSocket instead of `/run_sse`. Pass it on its own for a local autoboot, or alongside the agent's `https://` `--url` when deployed (for Agent Runtime, the full engine URL). The dataset, the trace output, and the `eval grade` step are unchanged.
# Local autoboot agents-cli eval generate --mode adk_live # Deployed agent agents-cli eval generate --mode adk_live --url https://my-live-agent.run.app --app-name app # Chain generate + grade agents-cli eval run --mode adk_live --url https://my-live-agent.run.app --app-name app \ --metrics final_response_quality
What gets graded
Live replies are audio, and are transcribed by default, so the **transcript** is what gets graded. Raw audio and video bytes are dropped from the trace.
How cases are played
Each user turn is sent as **text** over **one persistent socket**, in order. A single-turn case (`prompt`) is just a one-turn conversation. History lives in the live session rather than being seeded over HTTP, so the model conditions on the real running conversation, and multi-turn cases produce a trajectory that `multi_turn_*` metrics can score.
Author **user-only** turns. The agent generates every reply over the live session, so pre-authored agent turns are ignored (the CLI warns) and are not seeded as history.
Two failures that look like transport bugs
Both let the socket connect and then fail inside the session, so the error points at `/run_live` when the cause is agent configuration.
| Cause | Fix | |---|---| | The agent is not on a Live model. `--mode adk_live` changes only the transport; it cannot make a non-Live agent bidi. The scaffold default is not a Live model, so a fresh project fails with WebSocket code 1011. | Switch the agent to a Live model. | | On Vertex, the model's region is not pinned. Live models are served from a regional endpoint such as `us-central1`, not `global`. Unpinned, the model falls back to `GOOGLE_CLOUD_LOCATION`, which `agents-cli deploy` sets to `global` on Agent Runtime. | Pin the region on the model itself rather than steering `GOOGLE_CLOUD_LOCATION`, which is shared with sessions, telemetry, and grading. |
Grading is unaffected by either: `eval grade` passes its own location (`--region`, default `global`), so `eval run --mode adk_live` chains both steps correctly in one command.
> **ADK projects.** Current Live model IDs, and the code to pin a model's > region, are in `/google-agents-cli-adk-code` (`references/adk-python-live.md`, > "Models"). Transport mechanics are on the same page under "Events" and > "Serving and testing a Live agent".
Ships withgoogle-agents-cli
The CLI and skills that turn any coding assistant into an expert at creating, evaluating, and deploying AI agents on Google Cloud.
Get the whole plugin
Stats
6,002
Stars
678
Forks
Active
Maintenance
Python
Language
Apache-2.0
License
3d ago
Last commit
5mo ago
Created
Repo: google/agents-cli
Other agents on google-agents-cli.
@googleView AgentAgent
adk-go
Reflects `google.golang.org/adk/v2 v2.1.0`, the version the `adk_go` template pins. If a symbol here is missing, check your `go.mod` before assuming the page…
@googleView AgentAgent
adk-python-live
Real-time voice and video agents built on ADK's **Gemini Live API Toolkit** (`Runner.run_live` + `LiveRequestQueue`). Use this for low-latency spoken…
@googleView AgentAgent
adk-python-workflows
Requires `google-adk >= 2.0.0`. This page documents the Python graph API; ADK Go has its own — see `references/adk-go-workflows.md`. Requires **Python >=…
@googleView AgentAgent
adk-python
* **`Agent`**: The core intelligent unit. Can be `LlmAgent` (LLM-driven) or `BaseAgent` (custom/workflow). * **`Tool`**: Callable function providing external…
@googleView Agent