A Big Boost in Productivity
Data engineering is changing fast. Earlier, setting up a Fabric Data Agent meant spending 30 to 60 minutes clicking through portals and doing repetitive manual work. With AgentForge, this has changed completely.
AgentForge brings Fabric Data Agent setup into VS Code, using natural language powered by the Model Context Protocol (MCP). What once took nearly an hour can now be done in just 2 minutes for most agents. Even complex agents that work with large repositories are ready in about 4 minutes.
Creating and managing data agents using the Fabric portal can be slow and tiring, especially when working with many tables and schemas. Common challenges include:
- Too many manual steps just to create an agent and connect it to a lakehouse
- Scrolling through long schema lists and selecting tables one by one
- Writing agent instructions manually without seeing schema details clearly
- Switching constantly between the portal, SQL views, and Git repos
- Debugging issues manually by copying SQL into notebooks to test responses
All of this slows developers down and increases the chances of errors.
Introducing AgentForge: A Faster, AI‑Driven Way
AgentForge is a VS Code Copilot plugin designed to remove all this friction. Using MCP, it gives Copilot around 20 smart tools to handle everything related to a Fabric Data Agent, such as:
- Creating an agent
- Selecting tables
- Generating instructions
- Testing accuracy
- Tuning and publishing
Everything happens inside VS Code using simple natural language, turning a long, broken process into one smooth conversation.
How AgentForge Works (In Simple Terms)
AgentForge is built with a clean structure:
MCP Server
Provides tools for creating, updating, testing, and validating agents.3 Smart Assistants
- Manager: Handles the full agent lifecycle
- Explorer: Reads metadata safely
- Tester: Checks accuracy and results
Skills (Reusable Playbooks)
Step‑by‑step instructions like #create-agent, #test-agent, and #tune-agent that ensure best practices are always followed.
For example, when creating an agent, AgentForge automatically checks that the right tables are selected and verified—something that is easy to miss in manual workflows.
Real Example: ADOWIA Telemetry Agent
Here’s how a real telemetry agent setup looks with AgentForge:
- Create & connect agent to workspace – 30 seconds
- Select all required telemetry tables using plain language – 15 seconds
- Generate instructions automatically from Git or SQL views – 2 minutes
- Add example queries (few‑shots) with live SQL checks – 1 minute
- Publish and test the agent – 30 seconds
Even for this complex case, the entire flow takes about 4 minutes. For simpler agents, it’s often just 2 minutes from start to finish.
Built‑In Accuracy Testing and Auto‑Fixing
AgentForge makes sure agents stay accurate. Developers can upload a CSV file with expected questions and answers. The tool:
- Tests the agent automatically
- Compares results with expected values
- Allows small numeric differences (5% by default)
If something fails, AgentForge automatically runs this loop:
Reproduce → Diagnose → Fix → Publish → Re‑test
All of this happens inside the chat, without jumping between tools.
Fabric Portal vs AgentForge (Quick Comparison)
| Create agent | Clicking through UI | Simple natural language |
| Connect lakehouse | Multiple steps | One command |
| Select tables | Checkbox selection | Natural language (“select adowia tables”) |
| Write instructions | Manual writing | Auto‑generated from Git / SQL |
| Few‑shot examples | Not supported | Fully supported + SQL validated |
| Accuracy testing | Manual | Automated via CSV |
| Debug issues | Manual trial & error | Auto diagnose & fix loop |
| Git integration | Copy‑paste | Direct fetch from repo |
Getting Started
AgentForge makes building high‑quality Fabric Data Agents faster, simpler, and more reliable.
Sharing the steps below to use the plugin.
Step 1 — Install the plugin
Open GitHub Copilot Chat (Ctrl+Shift+I) or CLI and run:
Install Plugin from source harn_microsoft/fabric-copilot-plugins (Please reach out to me to get access on this repo)
The setup skill handles all prerequisites automatically.
Step 2 — Restart VS Code
Reload the window to activate the plugin.
Step 3 — Create your first agent
Open Copilot Chat and say: Create Data Agent
The #create-agent skill walks you through the full setup:
- Connect a lakehouse
- Select tables
- Auto-generate instructions from semantic models (TMDL files)
- Add few-shot examples with SQL validation
- Publish and test
Step 4 — Test accuracy
Create a CSV file with Number, Prompt, ExpectedAnswer columns, attach it in chat, and say:
Run Accuracy Test
Step 5 — Tune
If any test fails, just say: Tune the Agent
It reproduces the failure, diagnoses the root cause, applies the fix, re-publishes, and re-tests — all in one conversation.
Microsoft Employee