Amazon Bedrock Foundation models with Amazon Opensearch Serverless as a Vector DB
-
Updated
Oct 7, 2026 - Python
Amazon Bedrock Foundation models with Amazon Opensearch Serverless as a Vector DB
Building an Intelligent AWS Cloud Engineer Agent with Strands Agents SDK
A full-stack web application that conducts comprehensive research using multiple specialized AI agents to gather, analyze, and synthesize information from diverse sources into coherent reports.
The Audited Context Generation (ACG) Protocol prevents AI hallucinations with a dual-layer system. The UGVP layer links every fact to a precise source for verification. The RSVP layer audits the AI's logical reasoning when combining facts. This creates a fully transparent, machine-auditable trail for both source and logical integrity.
Composition root for production agent services — fork, customize, ship. Built on the foundry-agent-* packages.
Agentic AI assistant for proactive AWS Health Event management and impact analysis
It is an agent based on Strands SDK using MCP and SKILL.
From Douglas Hofstadter's "Gödel, Escher, Bach" (1979)
A comprehensive cookbook featuring 10 practical AI agent examples using Strands Agents SDK and Amazon Bedrock
Universal three-layer evaluation SDK for AI agents, with a reference deployment on Amazon Bedrock AgentCore and Strands.
Production-ready, AI-native agency OS with multi-tenant FastAPI backend, React frontend, Strands-style agent workflows, Sentinel policy scanning, JWT auth, RBAC, project/lead/invoice management, and AWS-ready Docker/Terraform deployment.
Reference architecture for secure multi-agent delegation chains using Cedar-based permission attenuation on Amazon Bedrock AgentCore. Demonstrates scope narrowing, tamper-proof token chains, and audit logging across A2A agent interactions.
A solver for Strands, the New York Times puzzle game.
AI co-pilot for insurance advisors built on Amazon Bedrock AgentCore — text + voice agents, MCP tools, and coverage-aware customer insight.
This package provides a Strands-native “cost guard” layer for AI agents, focused on controlling and attributing LLM and tool spend within Strands-based workflows. It enforces budgets and policies at run time while emitting persistent, OpenTelemetry-style metrics so teams can monitor and optimize cost across tenants, strands, and workflows.
To associate your repository with the strands topic, visit your repo's landing page and select "manage topics."