Time for round two with AgentCore!
This time I wanted to experiment with Amazon’s Strands package. To get into it I talked with Claude and learned it’s a fast-to-run framework, supports agent as tool and multi-agent, natural integration with AWS and can work without it too.
The summary: Let the LLM think, give it tools, let it run. Less suitable for processes with many manual conditions.
Building the Agent
After I understood I asked Claude Code to learn from the official docs and build itself a skill.
It also built an Orchestrator agent for demonstration, we connected it to Amazon Bedrock with Nvidia Nemotron 3 Nano model and it ran locally on my computer.
I wanted to add external tools to make the agent functional. After consulting with Claude we went with a simple start of search tools and instructions for trend research and summaries.
We also built a Telegram bot for the agent to talk through.
Deploy and Issues
After the agent was running Claude deployed to AgentCore. Some issues we overcame along the way:
- Dockerfile was missing - the agentcore CLI didn’t create it automatically, we created manually
- Incorrect Docker configuration caused rate limit error - was hard to diagnose because the error didn’t appear in CLI output, only in CloudWatch logs. We learned we need to pull them with Python
- Had to add secret manager to the cloud
- The agent couldn’t search - embedded observability in code so it could run in AgentCore
In the end the deploy succeeded and we updated Claude’s skill for next times. The skill from the previous deploy helped especially - the UI build settings that were complicated the first time.
The Result
Now the agent runs in AWS cloud and talks to me through the Telegram bot running locally on my machine. Can also upload the bot to Lambda so it won’t depend on running on my computer.
I also added permissions to send emails so it sends me scheduled updates when I request without depending on Telegram.
What did I do in all this? Just talked with Claude and configured a few things in AWS. It controls the terminal and if I understand what I want there’s no limit to what can be done.
Scheduled Tasks
The agent also executes scheduled tasks and reports to me via email! Searches, reports, reminders and whatever I want at the time I set for it.
How it works:
- Enabled Amazon SES for sending emails
- Configured AWS EventBridge
- Set up scheduler function in Lambda with prompt for agent “search AI news…”
- Connected EventBridge to function with
aws events put-targets - Function calls agent according to schedule
- Agent uses its tools and executes
- I receive email
Costs for Personal Use
Almost free:
- Lambda: free tier
- AgentCore: ~$0.01 per invocation
- SES: $0.10 per 1000 emails
- Secrets Manager: $0.40 per month per secret