Agent developers can now create an eval set and add eval cases through command line itself. Adding an eval case is limited only to specifying conversation scenarios.
Sample comamnds:
- Create an eval set:
adk eval_set create \
contributing/samples/hello_world \
set_01
- Add an eval case with scenario file
Content of scenarios.json file:
'{"scenarios": [{"starting_prompt": "hello", "conversation_plan": "world"}]}'
adk eval_set add_eval_case \
contributing/samples/hello_world \
set_01 \
--scenarios scenarios.json
PiperOrigin-RevId: 817456117
The `agent_loader.load_agent` method can now return an `App` object. This change unwraps the `App` to get its `root_agent` before passing it to the graph builder, makes sure a `BaseAgent` instance is always used
PiperOrigin-RevId: 817209601
Details:
- Introduces a concept of `ConversationScenario` to represent a scenario that user simulator is supposed to follow.
- Introduces a `UserSimulator` interface, that one should implement. UserSimulator interface will be integrated with LocalEvalService in subsequent PRs.
PiperOrigin-RevId: 816883699
When there are multiple intervals and compactions, the original implementation only keep the last one. The right implementation is to keep as many compaction events/summary as the requested internals.
PiperOrigin-RevId: 816516662
This plugin intercepts tool failures, provides structured guidance to the LLM for reflection and correction, and retries the operation up to a configurable limit.
**Key Features:**
- **Concurrency Safe:** Uses locking to safely handle parallel tool
executions
- **Configurable Scope:** Tracks failures per-invocation (default) or globally
using the `TrackingScope` enum.
- **Extensible Scoping:** The `_get_scope_key` method can be overridden to
implement custom tracking logic (e.g., per-user or per-session).
- **Granular Tracking:** Failure counts are tracked per-tool within the
defined scope. A success with one tool resets its counter without affecting
others.
- **Custom Error Extraction:** Supports detecting errors in normal tool
responses
that
don't throw exceptions, by overriding the `extract_error_from_result`
method.
**Example:**
```python
from my_project.plugins import ReflectAndRetryToolPlugin, TrackingScope
# Example 1: (MOST COMMON USAGE):
# Track failures only within the current agent invocation (default).
error_handling_plugin = ReflectAndRetryToolPlugin(max_retries=3)
# Example 2:
# Track failures globally across all turns and users.
global_error_handling_plugin = ReflectAndRetryToolPlugin(max_retries=5,
scope=TrackingScope.GLOBAL)
# Example 3:
# Retry on failures but do not throw exceptions.
error_handling_plugin =
ReflectAndRetryToolPlugin(max_retries=3,
throw_exception_if_retry_exceeded=False)
# Example 4:
# Track failures in successful tool responses that contain errors.
class CustomRetryPlugin(ReflectAndRetryToolPlugin):
async def extract_error_from_result(self, *, tool, tool_args,tool_context,
result):
# Detect error based on response content
if result.get('status') == 'error':
return result
return None # No error detected
error_handling_plugin = CustomRetryPlugin(max_retries=5)
```
PiperOrigin-RevId: 816456549
Merge https://github.com/google/adk-python/pull/2857
Adds support for invoking Gemma models via the Gemini API endpoint. To support agentic function, callbacks are added which can extract and transform function calls and responses into user and model messages in the history.
This change is intended to allow developers to explore the use of Gemma models for agentic purposes without requiring local deployment of the models. This should ease the burden of experimentation and testing for developers.
A basic "hello world" style agent example is provided to demonstrate proper functioning of Gemma 3 models inside an Agent container, using the dice roll + prime check framework of similar examples for other models.
## Testing
### Testing Plan
- add and run integration and unit tests
- manual run of example `multi_tool_agent` from quickstart using new `Gemma` model
- manual run of `hello_world_gemma` agent
### Automated Test Results:
| Test Command | Results |
|----------------|---------|
| pytest ./tests/unittests | 4386 passed, 2849 warnings in 58.43s |
| pytest ./tests/unittests/models/test_google_llm.py | 100 passed, 4 warnings in 5.83s |
| pytest ./tests/integration/models/test_google_llm.py | 5 passed, 2 warnings in 3.73s |
### Manual Testing
Here is a log of `multi_tool_agent` run with locally-built wheel and using Gemma model.
```
❯ adk run multi_tool_agent
Log setup complete: /var/folders/bg/_133c0ds2kb7cn699cpmmh_h0061bp/T/agents_log/agent.20250904_152617.log
To access latest log: tail -F /var/folders/bg/_133c0ds2kb7cn699cpmmh_h0061bp/T/agents_log/agent.latest.log
/Users/<redacted>/venvs/adk-quickstart/lib/python3.11/site-packages/google/adk/cli/cli.py:143: UserWarning: [EXPERIMENTAL] InMemoryCredentialService: This feature is experimental and may change or be removed in future versions without notice. It may introduce breaking changes at any time.
credential_service = InMemoryCredentialService()
/Users/<redacted>/venvs/adk-quickstart/lib/python3.11/site-packages/google/adk/auth/credential_service/in_memory_credential_service.py:33: UserWarning: [EXPERIMENTAL] BaseCredentialService: This feature is experimental and may change or be removed in future versions without notice. It may introduce breaking changes at any time.
super().__init__()
Running agent weather_time_agent, type exit to exit.
[user]: what's the weather like today?
[weather_time_agent]: Which city are you asking about?
[user]: new york
[weather_time_agent]: OK. The weather in New York is sunny with a temperature of 25 degrees Celsius (77 degrees Fahrenheit).
```
And here is a snippet of a log generated with DEBUG level logging of the `hello_world_gemma` sample. It demonstrates how function calls are extracted and inserted based on Gemma model interactions:
```
...
2025-09-04 15:32:41,708 - DEBUG - google_llm.py:138 -
LLM Request:
-----------------------------------------------------------
System Instruction:
None
-----------------------------------------------------------
Contents:
{"parts":[{"text":"\n You roll dice and answer questions about the outcome of the dice rolls.\n You can roll dice of different sizes...\n"}],"role":"user"}
{"parts":[{"text":"Hi, introduce yourself."}],"role":"user"}
{"parts":[{"text":"Hello! I am data_processing_agent, a hello world agent that can roll many-sided dice and check if numbers are prime. I'm ready to assist you with those tasks. Let's begin!\n\n\n\n"}],"role":"model"}
{"parts":[{"text":"Roll a die with 100 sides and check if it is prime"}],"role":"user"}
{"parts":[{"text":"{\"args\":{\"sides\":100},\"name\":\"roll_die\"}"}],"role":"model"}
{"parts":[{"text":"Invoking tool `roll_die` produced: `{\"result\": 82}`."}],"role":"user"}
{"parts":[{"text":"{\"args\":{\"nums\":[82]},\"name\":\"check_prime\"}"}],"role":"model"}
{"parts":[{"text":"Invoking tool `check_prime` produced: `{\"result\": \"No prime numbers found.\"}`."}],"role":"user"}
{"parts":[{"text":"The die roll was 82, and it is not a prime number.\n\n\n\n"}],"role":"model"}
{"parts":[{"text":"Roll it again."}],"role":"user"}
-----------------------------------------------------------
Functions:
-----------------------------------------------------------
2025-09-04 15:32:41,708 - INFO - models.py:8165 - AFC is enabled with max remote calls: 10.
2025-09-04 15:32:42,693 - INFO - google_llm.py:180 - Response received from the model.
2025-09-04 15:32:42,693 - DEBUG - google_llm.py:181 -
LLM Response:
-----------------------------------------------------------
Text:
{"args":{"sides":100},"name":"roll_die"}
-----------------------------------------------------------
...
```
COPYBARA_INTEGRATE_REVIEW=https://github.com/google/adk-python/pull/2857 from douglas-reid:add-gemma-via-api e6d015f6a9ccbcf20ef7a7af8e4bbe1e9a5936b6
PiperOrigin-RevId: 816451001
Ankur Sharma
Xinran (Sherry) Tang
Xiang (Sean) Zhou
Wei Sun (Jack)
George Weale
Kacper Jawoszek
Xuan Yang
Shangjie Chen
Google Team Member
Haoming Chen
Hangfei Lin
Che Liu
Douglas Reid