/google-antigravity-sdk

Customizing Retry & Backoff Behavior

By default, the agent backend automatically enables retries for transient errors:

• API Retries: Defaults to 2 retries (3 total attempts) with a 1,000ms initial exponential backoff for transient network errors (such as HTTP 429 rate limits, 5xx server errors, or connection drops).
• Model Output Retries: Defaults to 4 retries when the model produces malformed tool calls or output that fails structured schema validation (e.g., violating a configured response_schema).

This example demonstrates how to use RetryConfig to customize, override, or disable this default retry behavior without writing custom retry loops in your application code.

Code

from google.antigravity import Agent, LocalAgentConfig, types

# 1. Specialized Intent Presets
# Use built-in classmethod presets when running specialized workloads:
benchmark_config = LocalAgentConfig(
    retry_config=types.RetryConfig.benchmark()  # Unbounded API retries for 429/503 quota resilience
)

# 2. Advanced Explicit Configuration
# Use explicit Pydantic models to fine-tune both API retries and model output validation retries.
explicit_config = LocalAgentConfig(
    retry_config=types.RetryConfig(
        api_retry=types.ModelAPIRetryConfig(
            max_retries=5,
            initial_sleep_duration_ms=1000,
            exponential_multiplier=2.0,
            jitter_range=0.1,
        ),
        model_output_retry=types.ModelOutputRetryConfig(
            max_retries=2,  # Tighten output validation retries from default of 4 down to 2
        ),
    )
)

async with Agent(explicit_config) as agent:
    # The agent will use your custom retry thresholds and backoff multipliers
    # instead of the system defaults.
    response = await agent.chat("Summarize our latest system architecture.")
    print(await response.text())

Key Concepts

• Default Resilience: You do not need to configure retry_config for interactive chat; the backend automatically retries API errors and tool validation errors by default.
• RetryConfig.benchmark(): Unbounded preset designed for load tests and evaluation suites to survive transient 429/503 errors while inheriting production default model output retries.
• ModelAPIRetryConfig: Customizes network and API-level retries with exponential backoff and jitter. Ideal for high-flake environments or strict fail-fast requirements (max_retries=0).
• ModelOutputRetryConfig: Customizes how many times the backend prompts the model to correct malformed tool calls or schema validation mismatches before raising an error.