본문으로 건너뛰기

Local Testing Workflow

Learn the complete end-to-end workflow for testing agents locally using the FluxLoop CLI.

Workflow Overview​

The FluxLoop workflow follows these four main steps:

  1. Instrument: Add decorators to your agent code.
  2. Define: Set up personas and test inputs.
  3. Test: Run simulations and capture traces.
  4. Analyze: Upload results to the cloud for inspection.

1. Instrument Your Agent​

The first step is to mark your agent's entry point using the FluxLoop SDK. This allows FluxLoop to intercept inputs and outputs automatically.

# src/agent.py
import fluxloop

@fluxloop.agent(name="CustomerSupportAgent")
def my_agent(user_input: str) -> str:
    # Your agent logic (e.g., using LangChain, OpenAI, etc.)
    response = "Hello! How can I help you today?"
    return response

2. Define Test Scenarios​

FluxLoop uses a fluxloop.yaml file to manage test configurations. You can define Personas (who is testing) and Base Inputs (what they are asking).

# fluxloop.yaml
project:
  name: my-first-agent

runner:
  target: "src.agent:my_agent"
  type: python-function

personas:
  - name: novice_user
    description: A user new to the system, asks basic questions.
  - name: expert_user
    description: A power user who uses technical terminology.

base_inputs:
  - input: "How do I reset my password?"
    persona: novice_user
  - input: "What is the rate limit for the GraphQL API?"
    persona: expert_user

3. Generate Test Inputs​

Instead of testing only your base inputs, use FluxLoop's generation capabilities to create diverse variations. This helps surface edge cases in how your agent handles different phrasing.

# Generate 20 variations for each base input
fluxloop generate --limit 20

FluxLoop will use an LLM (configured in your .env or via fluxloop auth login) to create realistic test data.

4. Run Local Tests​

Execute your agent against the generated inputs:

fluxloop test

Options:​

  • --iterations 5: Run each input 5 times to test for non-determinism.
  • --parallel 4: Run 4 tests in parallel to speed up execution.
  • --skip-upload: Skip automatic upload after the test completes.

5. Analyze Results​

While results are saved locally in the results/ directory, the most effective way to analyze them is using the Web Platform.

# Upload the latest results (if you skipped auto upload)
fluxloop sync upload

Open the generated link to results.fluxloop.ai to:

  • Inspect Traces: View the step-by-step execution of your agent.
  • Compare Runs: See if your changes improved or degraded performance.
  • Collaborate: Share specific failure cases with your team.

6. Sync with the Cloud​

FluxLoop enables a "Cloud-First" workflow where scenarios are managed centrally.

# Pull team scenarios from the Web Platform
fluxloop sync pull

# Test against team scenarios
fluxloop test --scenario team-regression-v1

Next Steps​