Agentic Postgres Challenge Submission

This is a submission for the Agentic Postgres Challenge with Tiger Data

What I Built

I built Agentic Orchestrator, a full‑stack multi‑agent platform that showcases how AI agents can collaborate to manage and optimize a PostgreSQL database safely and autonomously. The project consists of a Next.js 14 front-end and a Fastify back-end, both written entirely in TypeScript. At its core is a workflow engine powered by LangGraph that coordinates five specialized agents: ETL, Search, Analyst, DBA, and Merge on top of Tiger Cloud’s Postgres service. Each agent performs a distinct function: the ETL agent forks the production database and loads data, the Search agent runs hybrid keyword + semantic searches, the Analyst agent extracts insights, the DBA agent applies optimizations in a fork, and the Merge agent promotes validated changes back to production. The idea was inspired by the challenge’s goal to encourage experimental, agent‑driven uses of Tiger’s zero‑copy forks and hybrid search capabilities

Demo

The complete project is available on GitHub at https://github.com/lumskii/agentic_orchestrator. The repository contains installation scripts, server and client code, as well as extensive documentation. For a quick overview, the project’s README includes a short video demo that shows an orchestrator running from fork creation through multi-agent execution and safe merge back to production. To reproduce the demo locally, clone the repository, install dependencies with’ npm run install:all’, configure your’ .env’ file with your Tiger and OpenAI keys, run the search-extension SQL script, and start the app with’ npm run dev’. You can then open the dashboard at http://localhost:3000 to create runs, view fork status, and test hybrid search. Of course, you will need to create you env variables.

How I Used Agentic Postgres

This project leverages several key features of Agentic Postgres:

Tiger CLI & MCP integration. The CLI is installed with a simple three‑command setup. It allows the application to authenticate with Tiger Cloud and programmatically create, list, and delete Postgres services and forks. MCP (Model Context Protocol) is installed alongside the CLI and enables natural‑language commands and access to Tiger documentation from within the agents. In our workflow, the ETL agent invokes the CLI to create a zero-copy fork and load data. The Merge agent uses it to promote the fork back into production, and the DBA agent can fetch best-practice guides via MCP.

Zero‑copy forks on Fluid Storage. Tiger’s copy‑on‑write forks provide an isolated environment for each agent run. The ETL agent instantly forks the production database, loads the working dataset, and hands it off to subsequent agents. This isolation protects production while allowing multiple agents to experiment in parallel. At the end of the workflow, the fork can be merged safely or discarded.

Hybrid search with pg_textsearch and pgvectorscale. The Search agent uses Tiger’s BM25 keyword search (pg_textsearch) for precise term matching and pgvectorscale for semantic similarity queries. It fuses the two methods using Reciprocal Rank Fusion (RRF) to return the most relevant results, as described in the README’s hybrid search section. The API exposes endpoints for BM25, vector, and hybrid searches, allowing users to compare these approaches.

Agentic workflow orchestration. Using LangGraph, the backend defines a multi‑step workflow where each agent runs in sequence and passes state to the next. Agents access Tiger Cloud via a tigerWrapper service that wraps CLI commands and an mcpClient for natural‑language queries. The modular design means you can add new agents (e.g., a summarizer or recommendation engine) by implementing a runXAgent() function and registering it in the orchestrator.

Overall Experience

Building with Agentic Postgres was a rewarding experience. The Tiger CLI installation was straightforward, and the ability to spin up instant forks felt revolutionary. Creating a branch of a 10 GB database in seconds and experimenting without fear of breaking production made iterative development enjoyable. Hybrid search proved surprisingly powerful: combining BM25 keyword ranking with vector similarity often surfaced documents that neither method could find alone. LangGraph made the multi-agent workflow declarative and easy to extend, and the integration of MCP promises exciting future capabilities, such as in-context documentation lookup. The biggest challenge was coordinating environment variables and encoding special characters in connection strings, but once configured, the system ran smoothly. Overall, Agentic Orchestrator demonstrates how Tiger’s agent‑native Postgres features can empower autonomous data agents to clean, search, analyze, and optimize databases safely and creatively.