Skip to content

AGENTS.md

Latest commit

 

History

History
545 lines (412 loc) · 27.7 KB

AGENTS.md

File metadata and controls

545 lines (412 loc) · 27.7 KB

AI Coding Assistant Context

This document provides context for AI coding assistants (Claude Code, Gemini CLI, GitHub Copilot, Cursor, etc.) to understand the ADK Python project and assist with development.

Project Overview

The Agent Development Kit (ADK) is an open-source, code-first Python toolkit for building, evaluating, and deploying sophisticated AI agents with flexibility and control. While optimized for Gemini and the Google ecosystem, ADK is model-agnostic, deployment-agnostic, and is built for compatibility with other frameworks. ADK was designed to make agent development feel more like software development, to make it easier for developers to create, deploy, and orchestrate agentic architectures that range from simple tasks to complex workflows.

Key Components

  • Agent - Blueprint defining identity, instructions, and tools (LlmAgent, LoopAgent, ParallelAgent, SequentialAgent, etc.)
  • Runner - Execution engine that orchestrates the "Reason-Act" loop, manages LLM calls, executes tools, and handles multi-agent coordination
  • Tool - Functions/capabilities agents can call (Python functions, OpenAPI specs, MCP tools, Google API tools)
  • Session - Conversation state management (in-memory, Vertex AI, Spanner-backed)
  • Memory - Long-term recall across sessions

Project Architecture

Please refer to ADK Project Overview and Architecture for details.

Source Structure

src/google/adk/
├── agents/          # Agent implementations (LlmAgent, LoopAgent, ParallelAgent, etc.)
├── runners.py       # Core Runner orchestration class
├── tools/           # Tool ecosystem (50+ files)
│   ├── google_api_tool/
│   ├── bigtable/, bigquery/, spanner/
│   ├── openapi_tool/
│   └── mcp_tool/    # Model Context Protocol
├── models/          # LLM integrations (Gemini, Anthropic, LiteLLM)
├── sessions/        # Session management (in-memory, Vertex AI, Spanner)
├── memory/          # Long-term memory services
├── evaluation/      # Evaluation framework (47 files)
├── cli/             # CLI tools and web UI
├── flows/           # Execution flow orchestration
├── a2a/             # Agent-to-Agent protocol
├── telemetry/       # Observability and tracing
└── utils/           # Utility functions

Test Structure

tests/
├── unittests/       # 2600+ unit tests across 236+ files
│   ├── agents/
│   ├── tools/
│   ├── models/
│   ├── evaluation/
│   ├── a2a/
│   └── ...
└── integration/     # Integration tests

ADK Live (Bidi-streaming)

  • ADK live feature can be accessed from runner.run_live(...) and corresponding FAST api endpoint.
  • ADK live feature is built on top of Gemini Live API. We integrate Gemini Live API through GenAI SDK.
  • ADK live related configs are in run_config.py.
  • ADK live under multi-agent scenario: we convert the audio into text. This text will be passed to next agent as context.
  • Most logics are in base_llm_flow.py and gemini_llm_connection.py.
  • Input transcription and output transcription should be added to session as Event.
  • User audio or model audio should be saved into artifacts with a reference in Event to it.
  • Tests are in tests/unittests/streaming.

Agent Structure Convention (Required)

All agent directories must follow this structure:

my_agent/
├── __init__.py      # MUST contain: from . import agent
└── agent.py         # MUST define: root_agent = Agent(...) OR app = App(...)

Choose one pattern based on your needs:

Option 1 - Simple Agent (for basic agents without plugins):

from google.adk.agents import Agent
from google.adk.tools import google_search

root_agent = Agent(
    name="search_assistant",
    model="gemini-2.5-flash",
    instruction="You are a helpful assistant.",
    description="An assistant that can search the web.",
    tools=[google_search]
)

Option 2 - App Pattern (when you need plugins, event compaction, custom configuration):

from google.adk import Agent
from google.adk.apps import App
from google.adk.plugins import ContextFilterPlugin

root_agent = Agent(
    name="my_agent",
    model="gemini-2.5-flash",
    instruction="You are a helpful assistant.",
    tools=[...],
)

app = App(
    name="my_app",
    root_agent=root_agent,
    plugins=[
        ContextFilterPlugin(num_invocations_to_keep=3),
    ],
)

Rationale: This structure allows the ADK CLI (adk web, adk run, etc.) to automatically discover and load agents without additional configuration.

Development Setup

Requirements

Minimum requirements:

  • Python 3.10+ (Python 3.11+ strongly recommended for best performance)
  • uv package manager (required - faster than pip/venv)

Install uv if not already installed:

curl -LsSf https://astral.sh/uv/install.sh | sh

Setup Instructions

Standard setup for development:

# Create virtual environment with Python 3.11
uv venv --python "python3.11" ".venv"
source .venv/bin/activate

# Install all dependencies for development
uv sync --all-extras

Minimal setup for testing only (matches CI):

uv sync --extra test --extra eval --extra a2a

Virtual Environment Usage (Required):

  • Always use .venv/bin/python or .venv/bin/pytest directly
  • Or activate with source .venv/bin/activate before running commands
  • Never use python -m venv - always create with uv venv if missing

Rationale: uv is significantly faster and ensures consistent dependency resolution across the team.

Building

# Build wheel
uv build

# Install local build for testing
pip install dist/google_adk-<version>-py3-none-any.whl

Running Agents Locally

For interactive development and debugging:

# Launch web UI (recommended for development)
adk web path/to/agents_dir

For CLI-based testing:

# Interactive CLI (prompts for user input)
adk run path/to/my_agent

For API/production mode:

# Start FastAPI server
adk api_server path/to/agents_dir

For running evaluations:

# Run evaluation set against agent
adk eval path/to/my_agent path/to/eval_set.json

ADK: Style Guides

Python Style Guide

The project follows the Google Python Style Guide. Key conventions are enforced using pylint with the provided pylintrc configuration file. Here are some of the key style points:

  • Indentation: 2 spaces.
  • Line Length: Maximum 80 characters.
  • Naming Conventions:
    • function_and_variable_names: snake_case
    • ClassNames: CamelCase
    • CONSTANTS: UPPERCASE_SNAKE_CASE
  • Docstrings: Required for all public modules, functions, classes, and methods.
  • Imports: Organized and sorted.
  • Error Handling: Specific exceptions should be caught, not general ones like Exception.

Autoformat (Required Before Committing)

Always run before committing code:

./autoformat.sh

Manual formatting (if needed):

# Format imports
isort src/ tests/ contributing/

# Format code style
pyink --config pyproject.toml src/ tests/ contributing/

Check formatting without making changes:

pyink --check --diff --config pyproject.toml src/
isort --check src/

Formatting Standards (Enforced by CI):

  • Formatter: pyink (Google-style Python formatter)
  • Line length: 80 characters maximum
  • Indentation: 2 spaces (never tabs)
  • Import sorter: isort with Google profile
  • Linter: pylint with Google Python Style Guide

Rationale: Consistent formatting eliminates style debates and makes code reviews focus on logic rather than style.

In ADK source

Below styles applies to the ADK source code (under src/ folder of the GitHub repo).

Use relative imports (Required)

# DO - Use relative imports
from ..agents.llm_agent import LlmAgent

# DON'T - No absolute imports
from google.adk.agents.llm_agent import LlmAgent

Rationale: Relative imports make the code more maintainable and avoid circular import issues in large codebases.

Import from module, not from __init__.py (Required)

# DO - Import directly from module
from ..agents.llm_agent import LlmAgent

# DON'T - Import from __init__.py
from ..agents import LlmAgent

Rationale: Direct module imports make dependencies explicit and improve IDE navigation and refactoring.

Always do from __future__ import annotations (Required)

Rule: Every source file must include from __future__ import annotations immediately after the license header, before any other imports.

# Copyright 2025 Google LLC
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

from __future__ import annotations  # REQUIRED - Always include this

# ... rest of imports ...

Rationale: This enables forward-referencing classes without quotes, improving code readability and type hint support (PEP 563).

In ADK tests

Use absolute imports (Required)

Rule: Test code must use absolute imports (google.adk.*) to match how users import ADK.

# DO - Use absolute imports
from google.adk.agents.llm_agent import LlmAgent

# DON'T - No relative imports in tests
from ..agents.llm_agent import LlmAgent

Rationale: Tests should exercise the same import paths that users will use, catching issues with the public API.

ADK: Local Testing

Unit Tests

Quick start: Run all tests with:

pytest tests/unittests

Recommended: Match CI configuration before submitting PRs:

uv sync --extra test --extra eval --extra a2a && pytest tests/unittests

Additional options:

# Run tests in parallel for faster execution
pytest tests/unittests -n auto

# Run a specific test file during development
pytest tests/unittests/agents/test_llm_agent.py

Testing Philosophy

Use real code over mocks: ADK tests should use real implementations as much as possible instead of mocking. Only mock external dependencies like network calls or cloud services.

Test interface behavior, not implementation details: Tests should verify that the public API behaves correctly, not how it's implemented internally. This makes tests resilient to refactoring and ensures the contract with users remains intact.

Test Requirements:

  • Fast and isolated tests where possible
  • Use real ADK components; mock only external dependencies (LLM APIs, cloud services, etc.)
  • Focus on testing public interfaces and behavior, not internal implementation
  • Descriptive test names that explain what behavior is being tested
  • High coverage for new features, edge cases, and error conditions
  • Location: tests/unittests/ following source structure

Docstring and comments

Comments - Explaining the Why, Not the What

Philosophy: Well-written code should be largely self-documenting. Comments serve a different purpose: they should explain the complex algorithms, non-obvious business logic, or the rationale behind a particular implementation choice—the things the code cannot express on its own. Avoid comments that merely restate what the code does (e.g., # increment i above i += 1).

Style: Comments should be written as complete sentences. Block comments must begin with a # followed by a single space.

Versioning

ADK adherence to Semantic Versioning 2.0.0

Core Principle: The adk-python project strictly adheres to the Semantic Versioning 2.0.0 specification. All release versions will follow the MAJOR.MINOR.PATCH format.

Breaking Change

A breaking change is any modification that introduces backward-incompatible changes to the public API. In the context of the ADK, this means a change that could force a developer using the framework to alter their existing code to upgrade to the new version. The public API is not limited to just the Python function and class signatures; it also encompasses data schemas for stored information (like evaluation datasets), the command-line interface (CLI), and the data format used for server communications.

Public API Surface Definition

The "public API" of ADK is a broad contract that extends beyond its Python function signatures. A breaking change in any of the following areas can disrupt user workflows and the wider ecosystem of agents and tools built with ADK. The analysis of the breaking changes introduced in v1.0.0 demonstrates the expansive nature of this contract. For the purposes of versioning, the ADK Public API Surface is defined as:

  • All public classes, methods, and functions in the google.adk namespace.

  • The names, required parameters, and expected behavior of all built-in Tools (e.g., google_search, BuiltInCodeExecutor).

  • The structure and schema of persisted data, including Session data, Memory, and Evaluation datasets.

  • The JSON request/response format of the ADK API server(FastAPI server) used by adk web, including field casing conventions.

  • The command-line interface (CLI) commands, arguments, and flags (e.g., adk deploy).

  • The expected file structure for agent definitions that are loaded by the framework (e.g., the agent.py convention).

Checklist for Breaking Changes:

The following changes are considered breaking and necessitate a MAJOR version bump.

  • API Signature Change: Renaming, removing, or altering the required parameters of any public class, method, or function (e.g., the removal of the list_events method from BaseSessionService).

  • Architectural Shift: A fundamental change to a core component's behavior (e.g., making all service methods async, which requires consumers to use await).

  • Data Schema Change: A non-additive change to a persisted data schema that renders old data unreadable or invalid (e.g., the redesign of the MemoryService and evaluation dataset schemas).

  • Tool Interface Change: Renaming a built-in tool, changing its required parameters, or altering its fundamental purpose (e.g., replacing BuiltInCodeExecutionTool with BuiltInCodeExecutor and moving it from the tools parameter to the code_executor parameter of an Agent).

  • Configuration Change: Altering the required structure of configuration files or agent definition files that the framework loads (e.g., the simplification of the agent.py structure for MCPToolset).

  • Wire Format Change: Modifying the data format for API server interactions (e.g., the switch from snake_case to camelCase for all JSON payloads).

  • Dependency Removal: Removing support for a previously integrated third-party library or tool type.

Commit Message Format (Required)

All commits must follow Conventional Commits format.

Format:

<type>(<scope>): <description>

[optional body]

[optional footer]

Common types: feat, fix, refactor, docs, test, chore

Examples:

feat(agents): Add support for App pattern with plugins

fix(sessions): Prevent memory leak in session cleanup

refactor(tools): Unify environment variable enabled checks

Rationale: Conventional commits enable automated changelog generation and version management.

Key Files and Locations

Quick reference to important project files:

  • Main config: pyproject.toml (uses flit_core build backend)
  • Dependencies: uv.lock (managed by uv)
  • Linting: pylintrc (Google Python Style Guide)
  • Auto-format: autoformat.sh (runs isort + pyink)
  • CLI entry point: src/google/adk/cli/cli_tools_click.py
  • Web UI backend: src/google/adk/cli/adk_web_server.py
  • Main exports: src/google/adk/__init__.py (exports Agent, Runner)
  • Examples: contributing/samples/ (100+ agent implementations)

Additional Resources

Python Tips

General Python Best Practices

  • Constants: Use immutable global constant collections (tuple, frozenset, immutabledict) to avoid hard-to-find bugs. Prefer constants over wild string/int literals, especially for dictionary keys, pathnames, and enums.
  • Naming: Name mappings like value_by_key to enhance readability in lookups (e.g., item = item_by_id[id]).
  • Readability: Use f-strings for concise string formatting, but use lazy-evaluated %-based templates for logging. Use repr() or pprint.pformat() for human-readable debug messages. Use _ as a separator in numeric literals to improve readability.
  • Comprehensions: Use list, set, and dict comprehensions for building collections concisely.
  • Iteration: Iterate directly over containers without indices. Use enumerate() when you need the index, dict.items() for keys and values, and zip() for parallel iteration.
  • Built-ins: Leverage built-in functions like all(), any(), reversed(), sum(), etc., to write more concise and efficient code.
  • Flattening Lists: Use itertools.chain.from_iterable() to flatten a list of lists efficiently without unnecessary copying.
  • String Methods: Use startswith() and endswith() with a tuple of strings to check for multiple prefixes or suffixes at once.
  • Decorators: Use decorators to add common functionality (like logging, timing, caching) to functions without modifying their core logic. Use functools.wraps() to preserve the original function's metadata.
  • Context Managers: Use with statements and context managers (from contextlib or custom classes with __enter__/__exit__) to ensure resources are properly initialized and torn down, even in the presence of exceptions.
  • Else Clauses: Utilize the else clause in try/except blocks (runs if no exception), and in for/while loops (runs if the loop completes without a break) to write more expressive and less error-prone code.
  • Single Assignment: Prefer single-assignment form (assign to a variable once) over assign-and-mutate to reduce bugs and improve readability. Use conditional expressions where appropriate.
  • Equality vs. Identity: Use is or is not for singleton comparisons (e.g., None, True, False). Use == for value comparison.
  • Object Comparisons: When implementing custom classes, be careful with __eq__. Return NotImplemented for unhandled types. Consider edge cases like subclasses and hashing. Prefer using attrs or dataclasses to handle this automatically.
  • Hashing: If objects are equal, their hashes must be equal. Ensure attributes used in __hash__ are immutable. Disable hashing with __hash__ = None if custom __eq__ is implemented without a proper __hash__.
  • __init__() vs. __new__(): __new__() creates the object, __init__() initializes it. For immutable types, modifications must happen in __new__().
  • Default Arguments: NEVER use mutable default arguments. Use None as a sentinel value instead.
  • __add__() vs. __iadd__(): x += y (in-place add) can modify the object in-place if __iadd__ is implemented (like for lists), while x = x + y creates a new object. This matters when multiple variables reference the same object.
  • Properties: Use @property to create getters and setters only when needed, maintaining a simple attribute access syntax. Avoid properties for computationally expensive operations or those that can fail.
  • Modules for Namespacing: Use modules as the primary mechanism for grouping and namespacing code elements, not classes. Avoid @staticmethod and methods that don't use self.
  • Argument Passing: Python is call-by-value, where the values are object references (pointers). Assignment binds a name to an object. Modifying a mutable object through one name affects all names bound to it.
  • Keyword/Positional Arguments: Use * to force keyword-only arguments and / to force positional-only arguments. This can prevent argument transposition errors and make APIs clearer, especially for functions with multiple arguments of the same type.
  • Type Hinting: Annotate code with types to improve readability, debuggability, and maintainability. Use abstract types from collections.abc for container annotations (e.g., Sequence, Mapping, Iterable). Annotate return values, including None. Choose the most appropriate abstract type for function arguments and return types.
  • NewType: Use typing.NewType to create distinct types from primitives (like int or str) to prevent argument transposition and improve type safety.
  • __repr__() vs. __str__(): Implement __repr__() for unambiguous, developer-focused string representations, ideally evaluable. Implement __str__() for human-readable output. __str__() defaults to __repr__().
  • F-string Debug: Use f"{expr=}" for concise debug printing, showing both the expression and its value.

Libraries and Tools

  • collections.Counter: Use for efficiently counting hashable objects in an iterable.
  • collections.defaultdict: Useful for avoiding key checks when initializing dictionary values, e.g., appending to lists.
  • heapq: Use heapq.nlargest() and heapq.nsmallest() for efficiently finding the top/bottom N items. Use heapq.merge() to merge multiple sorted iterables.
  • attrs / dataclasses: Use these libraries to easily define simple classes with boilerplate methods like __init__, __repr__, __eq__, etc., automatically generated.
  • NumPy: Use NumPy for efficient array computing, element-wise operations, math functions, filtering, and aggregations on numerical data.
  • Pandas: When constructing DataFrames row by row, append to a list of dicts and call pd.DataFrame() once to avoid inefficient copying. Use TypedDict or dataclasses for intermediate row data.
  • Flags: Use libraries like argparse or click for command-line flag parsing. Access flag values in a type-safe manner.
  • Serialization: For cross-language serialization, consider JSON (built-in), Protocol Buffers, or msgpack. For Python serialization with validation, use pydantic for runtime validation and automatic (de)serialization, or cattrs for performance-focused (de)serialization with dataclasses or attrs.
  • Regular Expressions: Use re.VERBOSE to make complex regexes more readable with whitespace and comments. Choose the right method (re.search, re.fullmatch). Avoid regexes for simple string checks (in, startswith, endswith). Compile regexes used multiple times with re.compile().
  • Caching: Use functools.lru_cache with care. Prefer immutable return types. Be cautious when memoizing methods, as it can lead to memory leaks if the instance is part of the cache key; consider functools.cached_property.
  • Pickle: Avoid using pickle due to security risks and compatibility issues. Prefer JSON, Protocol Buffers, or msgpack for serialization.
  • Multiprocessing: Be aware of potential issues with multiprocessing on some platforms, especially concerning fork. Consider alternatives like threads (concurrent.futures.ThreadPoolExecutor) or asyncio for I/O-bound tasks.
  • Debugging: Use IPython.embed() or pdb.set_trace() to drop into an interactive shell for debugging. Use visual debuggers if available. Log with context, including inputs and exception info using logging.exception() or exc_info=True.
  • Property-Based Testing & Fuzzing: Use hypothesis for property-based testing that generates test cases automatically. For coverage-guided fuzzing, consider atheris or python-afl.

Testing

  • Assertions: Use pytest's native assert statements with informative expressions. Pytest automatically provides detailed failure messages showing the values involved. Add custom messages with assert condition, "helpful message" when the expression alone isn't clear.
  • Custom Assertions: Write reusable helper functions (not methods) for repeated complex checks. Use pytest.fail("message") to explicitly fail a test with a custom message.
  • Parameterized Tests: Use @pytest.mark.parametrize to reduce duplication when running the same test logic with different inputs. This is more idiomatic than the parameterized library.
  • Fixtures: Use pytest fixtures (with @pytest.fixture) for test setup, teardown, and dependency injection. Fixtures are cleaner than class-based setup methods and can be easily shared across tests.
  • Mocking: Use mock.create_autospec() with spec_set=True to create mocks that match the original object's interface, preventing typos and API mismatch issues. Use context managers (with mock.patch(...)) to manage mock lifecycles and ensure patches are stopped. Prefer injecting dependencies via fixtures over patching.
  • Asserting Mock Calls: Use mock.ANY and other matchers for partial argument matching when asserting mock calls (e.g., assert_called_once_with).
  • Temporary Files: Use pytest's tmp_path and tmp_path_factory fixtures for creating isolated and automatically cleaned-up temporary files/directories. These are preferred over the tempfile module in pytest tests.
  • Avoid Randomness: Do not use random number generators to create inputs for unit tests. This leads to flaky, hard-to-debug tests. Instead, use deterministic, easy-to-reason-about inputs that cover specific behaviors.
  • Test Invariants: Focus tests on the invariant behaviors of public APIs, not implementation details.
  • Test Organization: Prefer simple test functions over class-based tests unless you need to share fixtures across multiple test methods in a class. Use descriptive test names that explain the behavior being tested.

Error Handling

  • Re-raising Exceptions: Use a bare raise to re-raise the current exception, preserving the original stack trace. Use raise NewException from original_exception to chain exceptions, providing context. Use raise NewException from None to suppress the original exception's context.
  • Exception Messages: Always include a descriptive message when raising exceptions.
  • Converting Exceptions to Strings: str(e) can be uninformative. repr(e) is often better. For full details including tracebacks and chained exceptions, use functions from the traceback module (e.g., traceback.format_exception(e), traceback.format_exc()).
  • Terminating Programs: Use sys.exit() for expected terminations. Uncaught non-SystemExit exceptions should signal bugs. Avoid functions that cause immediate, unclean exits like os.abort().
  • Returning None: Be consistent. If a function can return a value, all paths should return a value (use return None explicitly). Bare return is only for early exit in conceptually void functions (annotated with -> None).