AI/TLDR

Google · 2026-09-10 · major

Google ADK 2.9.0 — agents fail over to a backup model automatically

Google ADK 2.9.0 adds FallbackModel, which switches an agent to a backup model when the primary one returns an error. The Python release also ships a LiveKit runner for voice and telephony agents and loads ADK 2.0 graph workflows from YAML.

google/adk-python repository page on GitHub

Agents keep running when a model errors, and ADK now speaks over the phone through LiveKit.

Key specs

GitHub stars21,511

Quick facts

MakerGoogle
Versionv2.9.0
LicenseApache-2.0
LanguagePython
Installpip install google-adk
What's newFallbackModel, LiveKit runner, YAML graph workflows, MCP SDK 2.x
Breaking changes5

What is it?

FallbackModel is the headline addition in Google ADK 2.9.0: point an agent at a backup model and it takes over automatically when the primary model returns an error. The release also adds a LiveKit runner for voice and telephony agents, and lets ADK 2.0 graph workflows be written as declarative YAML files instead of Python code.

How does it work?

The release mostly widens what an ADK agent can talk to and what it reports. MCP servers on SDK 2.x are supported alongside 1.x, GcsArtifactService gained get_authenticated_url and get_signed_url, agent transfers carry a transfer_reason, and tool-trajectory evaluation gained an ignore_args option. Telemetry grew as well, with skill-load tracking, script-execution instrumentation and OTLP log export.

Why does it matter?

Model failover is plumbing most teams build themselves before an agent reaches production, and having it in the toolkit removes that step. The LiveKit runner opens phone and voice channels without a second stack, and YAML workflows let a graph be reviewed or edited by people who do not write Python. Five breaking changes ride along, so this is not a drop-in upgrade for every project.

Who is it for?

teams running Google ADK agents in production

Frequently asked questions

Does upgrading to ADK 2.9.0 break existing agents?
Google ADK 2.9.0 ships five breaking changes. Failed workflow nodes now rerun when a workflow resumes instead of replaying as complete, so node bodies must be idempotent. GCS tool file access is restricted to a configured local_file_root, InMemorySessionService raises SessionNotFoundError for unknown sessions, nested workflow retry behaviour changes, and under MCP 2.x unknown server fields no longer reach callers.
When does FallbackModel switch to the backup model?
FallbackModel in Google ADK 2.9.0 triggers on errors from the primary model: when the primary model errors, the agent automatically fails over to the configured backup so the run keeps going. Google describes the goal as safeguarding agent uptime, not as routing for cost or quality.
Can Google ADK build phone and voice agents now?
Yes. Google ADK 2.9.0 adds a LiveKit runner, which connects ADK agents to LiveKit for voice and telephony. That lets teams build agents that hold spoken conversations or answer calls, alongside the text and tool-calling agents ADK already supported, without assembling a separate voice stack.
Does ADK 2.9.0 work with MCP SDK 2.x?
Google ADK 2.9.0 supports Model Context Protocol servers on both SDK 1.x and 2.x, and 1.x remains the default install. One caveat applies under MCP 2.x: unknown fields sent by a server no longer reach the caller. Projects staying on MCP SDK 1.x are unaffected by that change.

Try it

pip install google-adk

Sources · 2 outlets

Tags

  • tool
  • repo
  • google
  • adk
  • agent-development-kit
  • ai-agents
  • python
  • livekit
  • voice-agents
  • telephony
  • mcp
  • workflows
  • open-source

← All releases · Learn AI