跳转至

Evaluating AI Agents: A Production Blueprint with Strands and AgentCore

Ch04.632 Evaluating AI Agents: A Production Blueprint with Strands and AgentCore

📊 Level ⭐⭐ | 5.5KB | entities/evaluating-ai-agents-production-blueprint-strands-agentcore.md

Evaluating AI Agents: A Production Blueprint with Strands and AgentCore

Overview

AWS blog post (2026-07-23) by Amit Deol, Hin Yee Liu, and Ryan Cormack presenting a production evaluation blueprint for AI agents using the Strands Agents SDK (strands-agents-evals) for build-time testing and Amazon Bedrock AgentCore Evaluations for production monitoring. Uses Motorway's dealer stock search agent as a worked example.

Three-Layer Build-Time Assessment

The evaluation framework operates across three distinct layers with pass/fail thresholds:

Layer 1: Tool Usage (>95% threshold) — Validates correct tool selection and parameter passing. Uses deterministic code-based graders (ToolSelectionGrader, TrajectoryOrderGrader) to verify which tools were called and the call sequence. Example: "Diesel vehicles from £7,000 to £20,000" should use search_vehicles with typed filters.

Layer 2: Reasoning (>85% threshold) — Assesses logical decision-making using LLM-as-judge evaluators (HelpfulnessEvaluator, TrajectoryEvaluator from strands-agents-evals). Ensures the agent's reasoning holds together; agents arriving at the right response through illogical reasoning will fail unpredictably.

Layer 3: Output Quality (>90% threshold) — Measures response helpfulness, accuracy, and actionability using LLM-as-judge evaluation (OutputEvaluator, GoalSuccessRateEvaluator).

All three layers must pass before deployment.

strands-agents-evals Framework

The strands-agents-evals framework provides three primitives:

  • Experiment: A collection of test cases run against the agent
  • Case: Input query, expected output, and expected tool trajectory
  • Evaluator: Scoring logic (deterministic or LLM-based)

Three grader types: - Code-based deterministic (Layer 1): Fast, cheap, reproducible — measures tool selection, parameter passing, trajectory ordering - LLM-as-judge (Claude Sonnet 4.6) (Layers 2-3): Flexible but non-deterministic — measures reasoning quality, output helpfulness, goal success - Human review (Calibration): Expensive, used to calibrate LLM judge prompts — handles edge cases and safety

Handling Non-Determinism

The run_all_layers() function accepts a num_trials parameter. Two key metrics:

  • pass@k: likelihood of succeeding at least once in k attempts
  • pass^k (pass to the kth): probability of succeeding in k consecutive trials — more important for customer-facing agents

Multi-turn conversation testing uses ActorSimulator to generate realistic multi-turn interactions and InteractionsEvaluator to score context retention.

Production Monitoring with AgentCore Evaluations

Two complementary modes:

  • On-demand evaluation: analyzes specific agent interactions by selecting spans from CloudWatch logs — useful for debugging
  • Online evaluation: automatically samples live traffic (1-5% sampling recommended) with up to 10 evaluators

Built-in evaluators: Builtin.Helpfulness (TRACE), Builtin.GoalSuccessRate (SESSION), Builtin.ToolSelection (TOOL_CALL), Builtin.Correctness (TRACE). Custom evaluators (e.g., DataFreshnessEvaluator, SafetyGuardrailEvaluator, DealerDataScopingEvaluator) handle domain-specific constraints.

Deployment Pipeline with Evaluation Gates

Five-phase pipeline: Build-time evaluation → Staging validation (on-demand AgentCore) → Shadow mode (4h minimum, 2% deviation threshold) → A/B testing (5% traffic) → Production rollout (100% traffic with continuous online evaluation). Each phase has defined thresholds that block deployment on failure.

Results

After implementing the pipeline: Tool selection accuracy 87%→98%, Task completion rate 82%→96%, Context retention 71%→94%, Production incidents 12→2 per month, Mean time to detect from hours to minutes.

Source

AWS Machine Learning Blog


关联

原文存档 - 相关概念: Harness Engineering