The brief, line by line

Every requirement in the Capstone 2 brief (Agentic AI Workshop 2026, "Business Performance Analyst Agent"), and where to see that it's met. "Recording" means a recorded live AI run in app/recordings/. You can watch each one in the app's Replay mode or online. Checked on 1 October 2026.

Result: every requirement is met. The one optional item not used is the backup dataset.

Your mission

What your agent must do

What a self-correction looks like. The AI guesses a column name that doesn't exist, and run_sql sends back the error with a hint (real output from tools.py):

{
  "error": "SQL error: ... no such column: yr",
  "query": "SELECT territory, SUM(revenue) FROM sales_lines WHERE yr = 2024 GROUP BY 1",
  "hint": "Check names with get_schema(). This is SQLite: strftime(), LIMIT, ||. Fix the query and try again."
}

The AI then rewrites the query with the right column (year = 2024) and runs it again.

Tools to build

A typical run_sql call and the start of its result. The result id (q1) is what make_chart and run_python use to pick up the rows:

run_sql("SELECT territory, ROUND(SUM(revenue), 0) AS revenue
         FROM sales_lines WHERE year = 2024 GROUP BY 1 ORDER BY 2 DESC")

10 row(s) [q1]
Southwest  9,121,932
Canada     6,231,210
Northwest  6,018,433
...

How it works

The brief's flow is: question → plan the analysis → write query or code → run and self-correct → chart and explanation → weekly briefing. Met: every stage shows up as a step on the Tracks screen, and Developer view adds the arguments, SQL, timings and AI rounds for each one.

flowchart LR
  accTitle: The brief's flow, from question to briefing
  Q(["Question"]) --> P["Plan the<br/>analysis"] --> W["Write query<br/>or code"] --> R["Run it"]
  R -- "error: fix and retry,<br/>up to 3 times" --> W
  R --> C["Chart and<br/>explanation"] --> B(["Weekly<br/>briefing"])
  classDef red stroke:#c62828,stroke-width:2px,color:#b71c1c
  class R red

Your data

Guardrails (non-negotiable)

Read-only is enforced three times over, so a mistake in one check can't change the data:

flowchart LR
  accTitle: Three locks between the AI and the data
  Q["SQL the AI wrote"] --> L1{"Starts with<br/>SELECT or WITH?"}
  L1 -- yes --> L2{"One statement, no<br/>data-changing words?"}
  L2 -- yes --> L3[("Database file<br/>opened read-only")]
  L3 --> R["Rows, and the query<br/>kept for the answer"]
  L1 -- no --> X["Refused,<br/>with a hint"]
  L2 -- no --> X
  classDef red stroke:#c62828,stroke-width:2px,color:#b71c1c
  class X red

A delete sent straight to run_sql (as on the Tools screen) is refused before it reaches the database:

{
  "error": "Only SELECT (or WITH ... SELECT) queries are allowed. The database is read-only.",
  "query": "DELETE FROM SalesOrderHeader WHERE OrderDate < '2023-01-01'"
}

In the recorded track the AI doesn't even try: it answers "I can’t help delete or alter data; the database is read-only, and I’m only able to run SELECT queries." What the agent is told is missing from the data (tools.py, AW_DICTIONARY):

"not_in_data": ["returns or refunds", "competitors or their prices", "customer satisfaction",
                "marketing spend", "website traffic", "budgets or targets", "weather", "staff morale",
                "any future figures (no forecasts)"],

Try these first

Scorecard on the recordings: 7 of 7: the 4 questions above plus the 3 guardrail traps, each marked against an answer worked out from the database at marking time.

Stretch goals and live demo

Workshop schedule and judging

A learning project, not a product. AdventureWorks is Microsoft's public sample data about a made-up bicycle company. Made by Victor Saly, Akashdeep Nijjar, Manuel Verduzco Valenzuela, Mazen Ahmed, Buddhika Gamage, Nathan Fryatt, Michael Kampouridis and Malak Sheat (the team). Source of this page: docs/requirements.md.