Mastodawn

Our own Kent Chen wrote up the multi-agent architecture the team built for Recce's AI Data Review.

Single agent kept forgetting findings as PRs got complex. Fix: orchestrator + two specialists, each with its own 200k context window.

One subagent fetches full PR context via a single GitHub GraphQL MCP call (replaced 5-10 gh CLI round-trips). The other explores data through 6 Recce MCP tools: lineage_diff, schema_diff, row_count_diff, custom queries.

Show thread

Recce - Trust, Verify, Ship

Subagents return summaries to orchestrator, not raw payloads. Built with Claude Agent SDK and MCP.

https://blog.reccehq.com/designing-reliable-ai-agents-for-dbt-data-reviews

#dbt #DataEngineering #AI #MCP

Designing Reliable AI Agents for dbt Data Reviews

Code changes have AI review tools. Data changes don' - until now. Here's how we went from a single prompt to an AI agent that performs the first pass on data validation in every PR.