Test Bed

The Test Bed is a live testing environment that opens when you click Run on any workflow card in the Agentic Workflows tab. It connects directly to your AgentCore runtime and streams the full run output in real time.

The Outcome

Evidence that a workflow produces the output you expect, scored on quality and probed for safety, before it reaches a real user.

Before You Start

You need a workflow to run and agents to run it:

• Workflow plans exist in Agentic Workflows
• Agents are configured in Agentic System - deployment to AgentCore fires automatically once the spec is ready, and the Test Bed connects to that runtime
• Any knowledge the agents need is loaded and bound to them - see Data

Testing before the data layer is in place tells you about the agents' reasoning, not about the answers they will actually give.

Start Here

Click Run on the workflow closest to the core of your use case - the one whose deliverable is the reason you built this system. Watch the full run before evaluating anything. The agent activity stream shows which agent picked up each step and which tools it called, and that is where a misassigned responsibility becomes visible.

Running a Workflow

Click Run on any workflow card. A run form may appear asking for any inputs the workflow needs. Submit to start the run.

The run streams live - you can see each agent activating, the tools it calls, and the reasoning as it happens. Once the run completes, review the structured output summary and any documents the agents generated.

Reviewing Output

Once the run completes, you can check the structured output that was generated and the document the agents produced.

Evaluate

After a run completes, click Evaluate to score the output across multiple quality dimensions. Select up to 8 metrics from the dialog:

Response Quality

• Correctness - are the facts accurate?
• Helpfulness - is the response useful and actionable?
• Coherence - does it flow logically?
• Conciseness - is it brief but complete?
• Faithfulness - is it grounded in retrieved data or agent reasoning?
• Relevance - does it directly address the question?
• Instruction Following - were the agent's instructions respected?
• Refusal Detection - did it avoid answering when it should not?

Task Completion

• Goal Success - did the agent's actions achieve the intended goal?

Tool Usage

• Tool Selection - did the agent pick the right tool?
• Tool Parameters - did the agent structure tool inputs correctly?

Safety

• Harmfulness - does the output contain harmful content?
• Stereotyping - does the output make unfair generalisations?

Click Run evaluation to start. Results show a rating per metric with detailed explanations.

Safety & Stress Test

Click Safety & Stress Test to probe your agents with structured adversarial techniques. Five attack methods are available:

• Crescendo - gradually escalating prompts that inch toward a policy boundary
• PAIR - iterative prompt injection that refines attacks based on agent responses
• GOAT - goal-oriented adversarial testing
• BadLikertJudge - exploits evaluation bias to bypass safety layers
• SequentialBreak - sequences of individually benign requests that combine to elicit unsafe output

You can view the live logs as the test runs:

Once complete, the results show whether your agent held firm or was successfully attacked.

What Good Looks Like

• Every workflow you intend to deploy has completed at least one full run
• The generated documents have been read by someone who knows the domain, not just skimmed for formatting
• Evaluation has been run with metrics chosen for this use case - Faithfulness and Tool Selection matter more for a retrieval-heavy agent than Conciseness does
• A Safety & Stress Test has been run, and any successful attack has been answered with a guardrail or prompt change
• You re-ran the workflow after making changes, rather than assuming the fix worked

Where This Leads

• Agentic System - where you fix what the run exposed: wrong agent, wrong tool, wrong prompt, or missing guardrail.
• Deploy - once runs are clean and the stress test results are acceptable.
• Governance & Guardrails - if the stress test succeeded against your agent, revisit the boundaries before deploying.