12:25–12:40
Wrap-up and clean-up
First: check your Intake scores
Your online evaluation should have scored the last batch of Intake questions by now.
In the AgentCore console, open Evaluations, then the Evaluation configurations tab, and choose intake_live. Results are shown in CloudWatch: use View results in AgentCore Observability, or go to GenAI Observability, Bedrock AgentCore and find intake_agent.
You set up intake_live in the console, so its scores are in the console too: follow the Console tab. To score the latest sessions on demand from the terminal instead:
$WORKSHOP/scripts/eval-harness.sh intake_agent AmountCheck Builtin.Correctness
Look for AmountCheck passing on the Lucky Star Print answers. If nothing's there yet, check again before you leave, because the scores will keep arriving.
What you built this morning
| Scenario | Where the agent ran | AgentCore pieces | What changed when you fixed things |
|---|---|---|---|
| Kopi Run | AgentCore harness | Harness, Runtime, Gateway, Memory, Observability | One line of a tool schema |
| Tally | Your terminal | Gateway, Policy, Observability | A schema entry and two Cedar rules. No agent code |
| Intake | AgentCore harness | Harness, Gateway, Code Interpreter, Evaluations | The system prompt and one extra tool |
None of the fixes needed new agent code. Each one was a change to the configuration around the model: a schema, a rule, a prompt, a tool. AgentCore lets you make those changes, version them and measure the effect without rebuilding the agent.
Questions to take back to your team
- Which of your existing APIs would make useful agent tools, and what would their schemas need to say?
- Where is your agent allowed to spend money, change data or see someone else's data? Those are your first Policy rules.
- How would you know today if your agent got worse after a prompt change?
- If you already run an agent, which single AgentCore piece would help most without moving it?
Bring your answers to the one-to-one consults straight after the session. AWS solutions architects and JAM engineers are in the room to talk them through with you.
Clean up
Workshop account: there's nothing to do. It closes automatically after the event, and we'll tell you when.
Your own account: if you ran any of this in your own AWS account, remove everything so it doesn't keep costing you money.
Delete these in the AgentCore console: the harnesses kopi_agent and intake_agent; the gateways kopi-gw, tally-gw and intake-gw; the policy engine tally_policies; the evaluator AmountCheck; and the evaluation configuration intake_live. Then, in the code editor terminal, remove the trace log group and the sample APIs:
aws logs delete-log-group --log-group-name /aws/bedrock-agentcore/runtimes/tally-billing-agent &&
$WORKSHOP/scripts/destroy-backends.sh
for p in KopiRun TallyTools IntakeQA; do
(cd $WORKSHOP/$p && agentcore remove all -y && agentcore deploy -y)
done &&
aws logs delete-log-group --log-group-name /aws/bedrock-agentcore/runtimes/tally-billing-agent &&
$WORKSHOP/scripts/destroy-backends.sh
agentcore remove all clears each project and the deploy that follows deletes its resources in AWS. The online evaluation intake_live was made in the console, so delete it there: Evaluations, Evaluation configurations.
destroy-backends.sh removes the three sample APIs if you deployed them with deploy-backends.sh; it asks before deleting anything. Deleting a harness also deletes the managed memory created for it. Transaction Search stays on; if you turned it on just for today, switch it off under CloudWatch, Application Signals (APM), Transaction Search.
Take it further
- AgentCore harness, including export to code
- AgentCore CLI and its command reference
- Gateway targets: Lambda, OpenAPI, Smithy, API Gateway and MCP servers
- Policy in AgentCore and Cedar
- Observability for agents hosted outside AgentCore
- AgentCore Evaluations
- AgentCore pricing. There's no separate charge for the harness; you pay for the pieces it uses.
- Kiro, if you want an AI IDE to build your own version of any of this
Tell us how it went
It's five questions and takes about two minutes. We'll share the link at the end of the session.
Thanks for spending your morning with us.