Data Support Agent v2

Overview

The original Data Support Agent was an n8n-based categorisation workflow — useful for routing, but without a reasoning loop it could triage requests but not resolve them. v1 automated around 50% of incoming requests. v2 is a complete rebuild that has already demonstrated the ability to automatically resolve approximately 90% of queries end-to-end.

What I built

Claude on AWS Bedrock runs as the reasoning layer, with a tool set covering Snowflake data queries, Confluence documentation search, and Jira ticket creation. When a request comes in via Slack, the agent decides what information it needs, calls the appropriate tools, and either answers directly or raises a ticket on the user's behalf — routing to the right team automatically.

The platform is config-driven: new bots are provisioned by dropping in a config file and a system prompt. No code changes, no deployment pipeline changes. Langfuse provides full observability across every request, tool call, and response, feeding directly into improvement cycles.

Expanding the platform

The same platform powers Virgil, a general-access bot available across the entire organisation. Virgil can pick up requests from any Slack channel, search GitHub for relevant code and Confluence for relevant documentation, then close the loop by raising a GitHub issue with a linked Jira ticket ready for implementation.

Outcomes

Went from a single-purpose categorisation bot to a multi-agent platform resolving ~90% of data support requests autonomously — up from ~50% with v1. Any team can stand up a new agent without engineering involvement, and continuous observability means quality improves over time rather than degrading silently.