Taxbit Assist: Deploying a Grounded, Role-Aware AI Agent on Amazon Bedrock AgentCore

Digital-asset tax and accounting platform Taxbit shared the architectural implementation of Taxbit Assist, an AI assistant integrated directly into its product dashboard. Built on Amazon Bedrock AgentCore runtime using the Strands Agents SDK, the system answers compliance and product workflow questions without giving tax advice or reading direct customer tax data. The agent relies on managed runtime memory for multi-turn session isolation and restricts identity handling by binding verified user roles and organization IDs into tool closures before model invocation. Retrieval-augmented generation (RAG) is powered by Amazon Bedrock Knowledge Bases and Amazon S3 Vectors, using semantic chunking enriched by document headers and lead paragraphs to preserve context. Query-time pipelines retrieve 75 vector candidates, rerank them server-side via Cohere Rerank 3.5, and pass up to five deduplicated passages to foundation models while enforcing strict citation rules and PII screening through Amazon Bedrock Guardrails.

Key takeaways

• Taxbit Assist runs on Amazon Bedrock AgentCore runtime to automate agent loops, multi-turn memory state, and session isolation across enterprise compliance users.
• The agent operates with bounded tools where identity parameters are bound via server-side closures, preventing prompt injections from elevating permissions.
• The RAG pipeline prepends document titles and lead scopes to semantic chunks, ingesting published docs, code-derived Markdown, and curated references into Amazon S3 Vectors.
• Incoming queries undergo Amazon Bedrock Guardrails PII screening before over-fetching 75 candidates and reranking them server-side down to five passages for model grounding.