Amazon Bedrock AgentCore harness: a first look
I created an AgentCore harness in the AWS console with no docs, added a browser tool and memory, saved a memory from the CLI, and the agent read it back.
The main idea
In the video I explore Amazon Bedrock AgentCore with no docs and no recent tutorials. I wanted a free-flowing session to see where I land. Creating an agent in the console starts with a harness name. I think of the agent as the model plus the harness.
I chose the GPT-6.1 Terra model with the Bedrock API source, because Mantle only had the 5.6 Terra. I wrote a system prompt for drafting tweets, enabled memory with a new memory managed by the harness, and added the browser tool. The console also shows export to code instructions and invocation code with boto3.
The browser tool bills by GB hour, and memory bills for ingestion, retrieval and storage. The agent opened my website and took a screenshot, but I could not open the file. It took a long route, through a base64 version, to see the image on my machine.
For memory, auto memory had already saved six facts from my playground chat. I saved a memory from the AgentCore CLI: my favorite cloud is GCP, but I also use AWS. I then asked the agent for my favorite cloud, and it answered GCP. My own hook that blocks Bedrock calls also got in the way and I had to unblock it.
Key points
- Creating an agent in AgentCore starts with creating a harness. The agent is the model plus the harness.
- The model source can be Bedrock, or OpenAI or Gemini directly. I used GPT-6.1 Terra on Bedrock.
- The browser tool costs per GB hour. Memory costs for ingestion, retrieval and storage.
- The console has export to code and view invocation code, so you can call the agent with boto3.
- Auto memory saved six facts from my playground chat without any setup.
- I saved a memory from the AgentCore CLI, and the agent answered GCP when I asked my favorite cloud.
Chapters
- 0:25Creating a harness in the console
- 3:57Enabling memory and the browser tool
- 6:53Export to code and costs
- 11:07Browser tool takes a screenshot
- 20:44Saving a memory from the CLI
- 24:36Agent reads the memory back
Save my free seat →