Agents Optimize

optimize

Measure and improve your AgentCore agent’s quality through evaluation, monitoring, and observability.

When to use

• You want to know if your agent is giving good answers
• You want to set up continuous quality monitoring in production
• You want to add a quality gate to your CI/CD pipeline
• You want to understand agent behavior through logs, metrics, and traces
• You want to set up CloudWatch dashboards or X-Ray tracing

Do NOT use for:

• Debugging a specific broken agent (wrong answers, errors) → use agents-debug
• Production security hardening (IAM, auth) → use agents-harden

Input

$ARGUMENTS can be:

• An eval goal: “add a quality gate”, “set up monitoring”
• An observability goal: “set up CloudWatch dashboard”, “understand my traces”
• A specific evaluator: “llm-as-a-judge”, “code-based”
• Empty — the skill will guide based on project context

Process

Step 0: Verify CLI version

Run agentcore --version. This skill requires v0.9.0 or later.

Step 1: Read project context

Read agentcore/agentcore.json to understand existing evaluators, online eval configs, and agent setup.

If agentcore/agentcore.json is not found:

“This skill requires an AgentCore project. Use agents-get-started to create one.”

Step 2: Determine the workflow

Step 3: Follow the loaded reference

The reference file contains the full procedure. Follow it step by step.

Cross-references

• After setting up evals, suggest agents-harden for production readiness
• If eval results reveal agent issues, suggest agents-debug for root cause analysis
• If the developer needs to add capabilities first, suggest agents-build

Output

Depends on the workflow — see the loaded reference for specific outputs.

Quality criteria

• Evaluator configuration uses only valid CLI flags
• Online eval sampling rate is appropriate (not 100% in production without discussion)
• CI/CD quality gate has a clear pass/fail threshold
• Observability setup includes both tracing and logging
• The developer understands the eval data delay: ~10 seconds put-to-get, end-to-end — one ingestion step covers both trace reads and eval queries; there is no separate indexing wait