Delfos: a specialized agent squad for Power BI development

Professional Power BI development is not a single-dimension problem. A well-designed semantic model needs a coherent star schema, efficient DAX measures, a report layout that communicates, RLS security that doesn’t penalize performance, and a DevOps pipeline that sustains everything in production. When a generic AI assistant tries to cover all of this at once, it loses context between layers and produces disconnected recommendations.

Delfos proposes an alternative: instead of a single generalist assistant, a coordinated squad of specialized agents that work against your real model through Power BI’s MCP servers.

From awesome-copilot to Delfos: from catalog to orchestration

The base components of Delfos started as individual contributions to the github/awesome-copilot repository — the official community catalog for GitHub Copilot customizations. Four expert agents (Data Modeling, DAX, Performance, Visualization), five instruction sets, and four skills were published as the power-bi-development plugin.

The catalog serves its purpose: reusable pieces anyone can install. But working with standalone pieces has an inherent problem: there’s no coordination between them. A DAX expert might optimize a measure without knowing that the Data Modeling Expert already chose a storage mode that invalidates that optimization. A Visualization Expert might recommend 8 visuals per page without knowing the Performance Expert detected the model can’t handle more than 5 within acceptable response times.

Delfos solves this with two orchestration agents. The Delfos Architect designs the complete project architecture in a unified way — model, DAX, reports, security, DevOps, and performance — in a single architecture document that becomes the source of truth. The Delfos Lead Squad takes that document and orchestrates implementation in phases, delegating each phase to the corresponding expert and pausing the workflow with human-in-the-loop (HITL) approval gates at every boundary.

MCP: agents that work against the real model

What differentiates Delfos from a prompt collection is its integration with the two Power BI MCP servers that Microsoft provides.

The Remote MCP (hosted on Fabric) is the read path. It connects to any published semantic model, discovers its schema, and executes DAX queries. When the DAX Expert validates a measure or the Performance Expert diagnoses a slow query, they’re executing against real data — not guessing.

The Modeling MCP (local VS Code extension) is the write path. It exposes the full Tabular Object Model: tables, columns, measures, relationships, security roles, TMDL files. When the Data Modeling Expert implements a star schema or the Lead Squad orchestrates a multi-phase build, they’re modifying the model directly.

Each agent knows when to use which. The Architect inspects via Remote before designing. The Data Modeling Expert implements via Modeling. The DAX Expert uses both: creates measures with Modeling and validates them with Remote. The Visualization Expert discovers available measures via Remote before mapping visuals. Every Modeling MCP write operation includes a confirmation prompt — the protocol itself has HITL built in.

The Business Central bridge: from star schema to API Pages

In projects where Power BI consumes data from Business Central, there’s a step that has historically been manual and error-prone: determining which BC API Pages you need to feed your semantic model. Delfos includes a specific skill for this: BC Data Source Mapping.

The skill takes the architecture document produced by the Architect, maps each star schema table against a catalog of standard BC v2.0 APIs organized by functional domain (Sales, Purchasing, Master Data, Finance, Inventory), and classifies each table as covered, partially covered, or not covered.

For covered tables, it confirms which standard API to use and recommends a $select to optimize the Power BI connector query. For tables with gaps, it generates an ALDC-compatible spec — ALDC being the agentic AL development framework I use for Business Central projects. That spec feeds into ALDC as just another requirement that al-spec.create consumes and al-conductor implements.

The result is a complete pipeline: Delfos designs what data the model needs, the bridge skill identifies how to get it from BC, and ALDC builds the AL API Pages when standard ones fall short. No manual translation between frameworks.

The quickstart: a Sales Order Tracker in 30 minutes

To keep this practical rather than theoretical, Delfos includes a step-by-step quickstart that builds a complete sales order tracking dashboard using the full workflow.

The flow starts with the Architect designing the architecture (star schema with FactSalesOrder, DimCustomer, DimDate, DimOrderStatus), moves through the BC mapping skill that verifies the data sources, continues with the Lead Squad orchestrating five phases (model, DAX, RLS, reports, performance) with a different expert in each and a HITL gate between phases, and ends with a completion document and the project memory updated.

Each phase shows the exact prompt, the expected result, and which MCP server is used. The quickstart works with a sample CSV so anyone can run it without needing a live BC environment.

References

• Delfos repository on GitHub
• power-bi-development plugin on awesome-copilot
• Power BI MCP Servers — Microsoft Learn documentation
• Power BI Modeling MCP — GitHub repository
• BC API v2.0 — Microsoft Learn documentation
• ALDC — AL Development Collection for GitHub Copilot

Delfos doesn’t aim to replace the Power BI developer’s knowledge — it aims to structure it. Every design decision gets documented in an architecture.md. Every implementation phase passes through an approval gate. Every validation runs against the real model. The difference between «asking an AI for help» and «working with a squad of specialized agents» is precisely that: structure, shared context, and traceability.

The repository is open source under MIT license. Use the agents as-is, adapt the instructions to your stack, or contribute new experts.

The oracle is ready.

Original Post https://techspheredynamics.com/2026/03/31/delfos-a-specialized-agent-squad-for-power-bi-development/