Skip to main content
Atla isn’t just for monitoring production traces and shipping improvements faster. It’s also used to help teams iterate so they can launch in days, not weeks.

​
Development Workflow Steps

1

Create custom metrics

Navigate to your Metrics screen and create custom metrics that matter to your domain. While we automatically flag error types and failure patterns, you can add domain-specific metrics like tool_call_efficiency.
2

Configure metadata for test runs

Set up metadata to track different test configurations. For example, you might run three versions of a prompt that are concise, balanced, and verbose:
from atla_insights import configure, set_metadata

# Configure global metadata for the experiment
configure(
    token="<YOUR_TOKEN>",
    metadata={"experiment": "prompt_comparison"}
)

# Then set dynamic metadata for each test run
set_metadata({"prompt": "v1"})  # concise version
set_metadata({"prompt": "v2"})  # balanced version  
set_metadata({"prompt": "v3"})  # verbose version
3

Run your tests

Execute your test suite with the configured metadata and custom metrics enabled.
4

Compare results

Navigate to the Compare screen to analyze the relative error rates and performance on your custom metrics across different configurations.
5

Deep dive into issues

Click “View” on any column to deep dive into the specific step-level errors in the traces to understand deeper issues.
6

Continue iterating

Continue to iterate on different dimensions. Use the “experiment” metadata tag to track performance of different experiments, such as testing different architectures and setups.

​
Example Metadata Configuration

from atla_insights import configure, set_metadata

# Configure with global experiment metadata
configure(
    token="<YOUR_TOKEN>",
    metadata={"experiment": "model_comparison"}
)

# Set dynamic metadata during runtime
set_metadata({"prompt": "v1", "model": "gpt-4"})
set_metadata({"prompt": "v2", "model": "claude-3"})
This approach helps you systematically test and improve your AI applications before they reach production.