Agent Lifecycle: Beyond the "Run" Button

In a professional setting, an agent isn't just a script you run once. It follows a lifecycle similar to traditional software, but with a major twist: Model Drift. An agent that works today might start failing tomorrow because the underlying LLM was updated or the user behavior changed.

1. The 5 Stages of the Lifecycle

Develop: Writing the system prompt, defining tools, and initial testing.
Evaluate: Running the agent against a "Golden Dataset" (Module 5) to ensure a high success rate.
Deploy: Pushing the agent configuration to the production environment.
Monitor: Tracking token usage, latency, and "Correctness" in real-time.
Retrain/Refine: Updating the prompt or substituting the model based on monitor feedback.

2. Versioning Agents

You should never "Overwrite" your production agent. If you change a prompt, you create v1.1.

The Reason: If v1.1 starts hallucinating, you need a single button to "Roll back" to v1.0.

What to Version?

The System Prompt.
The Tool Definitions.
The Model ID (e.g., gpt-4-0613 -> gpt-4-turbo).

3. Visualizing the Continuous Loop

graph LR
    Draft[Draft v1] --> Test[Eval Benchmarks]
    Test -- Pass --> Prod[Production v1]
    Prod --> Log[Observe Results]
    Log --> Drift[Detection: Errors up 5%]
    Drift --> Refine[Refine v2]
    Refine --> Test

4. The "Champion-Challenger" Pattern

Before you fully replace an agent, run a Shadow Test.

90% of traffic goes to the Champion (v1.0).
10% of traffic goes to the Challenger (v1.1).
If the Challenger performs better, it becomes the new Champion.

5. Agent "Retirement" (Deprecation)

Eventually, an agent becomes obsolete.

Example: A "Holiday Sale Assistant" is no longer needed in January.
The ADK Rule: Don't just delete the code. Deprecate it by disabling its entry point in the registry so that any other scripts trying to call it receive a clear "Out of Service" error instead of a crash.

Key Takeaways

Version control is mandatory for prompts and model IDs.
Shadow deployments reduce the risk of rolling out bad agent logic.
Monitoring for drift is the only way to ensure long-term stability.
The lifecycle is an infinite loop, not a finish line.

Module 10 Lesson 2: Agent Lifecycle Management