Feature Flags

Feature flags allow you to control agent behavior, prompt versions, and experimental features at runtime without deploying new code. They are defined in services/backend/src/config/feature_flags.py.

Overview

The feature flag system uses a FeatureFlag enum with a simple boolean check function. Flags can be toggled through environment variables, the admin API, or the dashboard UI.

Using Feature Flags


from src.config.feature_flags import FeatureFlag, is_feature_enabled
 
# Check if a flag is enabled
if is_feature_enabled(FeatureFlag.USE_V3_STYLE):
    prompt = load_prompt("style_v3")
else:
    prompt = load_prompt("style_v2")

Available Flags

Feature flags are defined as an enum in the configuration module:


class FeatureFlag(str, Enum):
    USE_V3_STYLE = "use_v3_style"
    ENABLE_TRY_ON = "enable_try_on"
    ENABLE_SHOPPING_AGENT = "enable_shopping_agent"
    USE_DEEPSEEK_CRITIC = "use_deepseek_critic"
    ENABLE_MEMORY_V2 = "enable_memory_v2"
    ENABLE_COST_ALERTS = "enable_cost_alerts"
    ENABLE_APPROVAL_QUEUES = "enable_approval_queues"
    AB_TEST_PERSONALIZATION = "ab_test_personalization"

Controlling Agent Versions

Feature flags are commonly used to switch between agent prompt versions during rollouts:


from src.config.feature_flags import FeatureFlag, is_feature_enabled
 
class StyleAgent(BaseAgent):
    async def execute(self, input_data: dict):
        if is_feature_enabled(FeatureFlag.USE_V3_STYLE):
            system_prompt = STYLE_V3_PROMPT
            model = "gpt-5.1"
        else:
            system_prompt = STYLE_V2_PROMPT
            model = "gpt-4o"
 
        response = await self.llm.generate(
            model=model,
            system=system_prompt,
            messages=[{"role": "user", "content": input_data["description"]}],
        )
        return response

A/B Testing with Feature Flags

Feature flags integrate with the personalization A/B testing system to run controlled experiments:


from src.config.feature_flags import FeatureFlag, is_feature_enabled
from src.services.personalization_ab_testing import get_user_variant
 
async def get_recommendations(user_id: str, input_data: dict):
    if is_feature_enabled(FeatureFlag.AB_TEST_PERSONALIZATION):
        variant = get_user_variant(user_id, "personalization_v2")
 
        if variant == "treatment":
            return await personalized_recommendations_v2(user_id, input_data)
        else:
            return await personalized_recommendations_v1(user_id, input_data)
 
    return await personalized_recommendations_v1(user_id, input_data)

Users are consistently assigned to the same variant for the duration of the experiment based on a hash of their user ID and the experiment name.

Gradual Rollout Patterns

Percentage-Based Rollout

Roll out a new feature to an increasing percentage of users:


from src.config.feature_flags import is_feature_enabled_for_user
 
async def analyze(user_id: str, input_data: dict):
    # Enable for 25% of users based on user_id hash
    if is_feature_enabled_for_user(
        FeatureFlag.ENABLE_MEMORY_V2,
        user_id,
        rollout_percent=25,
    ):
        memory = MemoryServiceV2()
    else:
        memory = MemoryServiceV1()
 
    return await run_pipeline(input_data, memory=memory)

Organization-Based Rollout

Enable features for specific organizations (B2B):


from src.config.feature_flags import is_feature_enabled_for_org
 
async def run_workflow(org_id: str, workflow_id: str):
    if is_feature_enabled_for_org(
        FeatureFlag.ENABLE_APPROVAL_QUEUES,
        org_id,
    ):
        # Run with approval gates
        return await execute_with_approvals(workflow_id)
    else:
        return await execute_workflow(workflow_id)

Configuration Sources

Feature flags are resolved from the following sources in priority order:

Admin API / Dashboard UI — Highest priority. Set through the B2B admin endpoints.
Environment variables — Set in the .env file. Use the flag name in uppercase with a FF_ prefix:
```
FF_USE_V3_STYLE=true
FF_ENABLE_TRY_ON=false
```
Default values — Defined in feature_flags.py. Used when no override is set.

Viewing Flag Status

Check the current state of all feature flags through the B2B admin API:


GET /api/v1/admin/feature-flags


{
  "flags": [
    { "name": "use_v3_style", "enabled": true, "source": "env" },
    { "name": "enable_try_on", "enabled": false, "source": "default" },
    { "name": "enable_analysis_agent", "enabled": true, "source": "admin" }
  ]
}