Markdown source

Slides: Teaching Coding Agents to do Spreadsheets - Nuno Campos, Witan Labs

Source Video

Teaching Coding Agents to do Spreadsheets - Nuno Campos, Witan Labs

Relationship To World's Fair 2026

These slides are extracted from a public AI Engineer YouTube video connected to World's Fair 2026. Speaker-matched clips are supporting context unless later confirmed as exact session recordings; official livestream recordings are day-level/event-level source material.

Related Scheduled Sessions

• No individual scheduled session mapping has been assigned yet; treat this as an event livestream deck.

Extracted Slides

• Recreated text/layout view: open HTML recreation
• AI slide classifier: content_slide confidence 0.98
• Text source: agent_vision.

Slide text:

50% → 92%

- 4 months, multiple architectures, and many dead ends

- What mattered most: replacing 15 discrete tools with one REPL

• Recreated text/layout view: open HTML recreation
• AI slide classifier: content_slide confidence 0.98
• Text source: agent_vision.
• OCR decision: ready — dense table/small text and formula-like content

Slide text:

The problem

• Recreated text/layout view: open HTML recreation
• AI slide classifier: content_slide confidence 0.97
• Text source: agent_vision.

Slide text:

One dead end

Three specialized agents:

1. Block discovery – identifies workbook structure

2. Edit agent – 5-step process: disambiguate, define end state, plan, execute, verify

3. Question agent – answers questions

Key finding: Rigid architectures don't win

• Recreated text/layout view: open HTML recreation
• AI slide classifier: content_slide confidence 0.97
• Text source: agent_vision.
• OCR decision: ready — table layout with small text across multiple rows

Slide text:

More dead ends

• Recreated text/layout view: open HTML recreation
• AI slide classifier: content_slide confidence 0.96
• Text source: agent_vision.
• OCR decision: ready — code snippets and small before/after tool-call text

Slide text:

Before vs. After

• Recreated text/layout view: open HTML recreation
• AI slide classifier: content_slide confidence 0.95
• Text source: agent_vision.
• OCR decision: ready — paragraph text plus code snippet on slide

Slide text:

Code mode vs. REPL

• Recreated text/layout view: open HTML recreation
• AI slide classifier: content_slide confidence 0.94
• Text source: agent_vision.
• OCR decision: ready — diagram and small explanatory text are better suited to OCR

Slide text:

The verification loop

• Recreated text/layout view: open HTML recreation
• AI slide classifier: content_slide confidence 0.98
• Text source: agent_vision.

Slide text:

Interface vs. engines

The REPL is an interface — the best one today, because coding is where models are strongest.

The engines — formula calculation, rendering, linting — are the more durable part. They're what close the verification loop, and they compound with each new model.

If for instance agents become as capable at computer use as they are at coding, the interface might change. The engines won't.

• Recreated text/layout view: open HTML recreation
• AI slide classifier: content_slide confidence 0.99
• Text source: agent_vision.

Slide text:

Domain knowledge outlived every tool

— We changed tools four times in four months

— The financial domain knowledge improved results on every one of them

Structured as a composable prompt component:

— How to interpret margins, profitability, revenue cascades

— Model type recognition (DCF, LBO, three-statement)

— Communication conventions ($1.2M not $1,234,567.89)

— "Never calculate in your head what the spreadsheet can calculate for you"

It was the most reused component in the system.

• Recreated text/layout view: open HTML recreation
• AI slide classifier: content_slide confidence 0.99
• Text source: none.
• OCR decision: ready — Dense two-column table with small text is better handled by OCR.

Hidden Non-Slide Evidence

• `slide-001.jpg` — speaker_stage confidence 0.99; Speaker on stage with a cropped projected slide; not a readable presentation frame.

Classification audit: raw/sources/slide-ai-classification/slides/HEFSExa0xl0/audit.json

Slide-Derived Subjects To Review

Subject extraction uses video title, related session titles/descriptions, transcript context, and OCR text when available. OCR is best-effort and should be reviewed against the embedded slide images.