Platform overview

Polaris runs on AWS managed services. The agent runs on Amazon Bedrock AgentCore, and every model call goes to Claude models on Amazon Bedrock.

The pieces

Piece	What it does
Web app	The Polaris interface you use in the browser, served over HTTPS through Amazon CloudFront. It handles accounts, chats, stories and settings.
Agent runtime	Each chat gets its own agent session on AgentCore Runtime. The session holds the chat’s working state and runs the agent loop: reading context, running SQL, drawing charts. Separate sessions keep chats isolated from each other.
Models	Claude Opus, Sonnet and Haiku on Amazon Bedrock. See Models.
Code sandbox	AgentCore Code Interpreter runs Python and shell code that the agent writes, in an isolated sandbox.
Browser	AgentCore Browser opens web pages for web fetch in an isolated cloud browser.
Web search	A web search tool through AgentCore Gateway answers web search requests.
Data	The example data is a DuckDB database that Polaris opens read-only. Other databases are queried where they live.
Observability	Traces of agent runs and model calls go to Amazon CloudWatch.

What happens when you ask a question

1. The web app sends your message to the chat’s agent session on AgentCore Runtime. If the chat has no active session, one starts, which can take about a minute.
2. The agent reads the project context and calls Claude on Bedrock.
3. It runs tools as needed: SQL on your databases, the code sandbox, web search or the browser.
4. Each step streams back to your browser as it happens.
5. The answer, queries and charts are saved with the chat.

Security basics

• You sign in with an email and password. Each user sees only the projects and shared items they have access to.
• SQL runs read-only by default.
• Connections use HTTPS.
• Access to the web app is limited to the United States.
• Model usage is tagged and tracked, and admins can set budgets.