Making deck.gl
AI-Ready

Everybody uses AI to write code now.

So… how well does AI treat deck.gl?

Five questions, five discussions

1. What do models recommend when someone asks for a map?
2. How well do frontier models write deck.gl?
3. How can agents drive deck.gl interactively?
4. Does the website and documentation work for AI?
5. How do we handle AI contributions?

Given a prompt like "Make me a map that…"

A clear progression. Older models (GPT-4o, Haiku 4.5, GPT-5-mini) answered Leaflet for anything simple, Mapbox for React and deck.gl 8. The frontier answers MapLibre for the basemap, deck.gl for the data, and Leaflet only for a throwaway page.

Do they know today's deck.gl? No. None gets past 9.1.

We need one deck.gl + MapLibre story. Our MapLibre page offers three modes, two packages, two React roots. In 75 generated maps the models took four routes; 12 still carry a v8 shim. One path per environment, the rest retired in v10.

How we tested it

Seven models: Opus 4.8, Fable 5.1, GPT-5.6 Sol, GPT-6 Astra, Gemini 3.1 Pro, Gemini 3.8 Flash, GLM 5.3. Real public data, CDN scripts only, no API keys.

Almost every map works.

Score is the four-task total out of 68. Thumbnails are the 8-second screenshot of the generated page, unedited.

The open brief is where they differ

We have to update the examples

• The code is a generation behind. Most files load a deck.gl 8.9 bundle or use v8 patterns; only GPT-5.6 Sol used the current overlay integration; none used 9.2 or later.
• They copy the showcase examples. The hexagon and trips tasks pulled every model toward the official example's stack, palette and defaults, v8 included.
• Same cartographic choices, same blind spots. Every choropleth is YlOrRd quantiles; skewed data goes unhandled; legends vanish when not demanded. Good, never excellent.
• The open-weight model knows the words, not the rules. GLM 5.3 mutated layer props instead of replacing layers and shadowed the library global.

Many applications will use deck.gl with agents to drive dynamic maps

Not "write me an app" but "show me", "filter that", "color it by income", while the map is running. The most logical interface for that is one deck.gl has had since 2018: the agent writes declarative JSON.

We are lucky to have it: declarative, portable, no eval, and models already know its syntax. What follows is how far it gets us today, and what is missing.

Everyone is already doing it

"a growing need to generate powerful visualizations directly from the backend… without having knowledge about how to code front-end applications."Ib Green, RFC: JSON Layers, July 2018. Replace "backend" with "agent". The v2 tracker opened in April 2026 names LLMs as goal #1.

How CARTO does it, two ways

Both work in production. Both hit the same wall.

We flood the agent's context, and it never learns what happened

An agent recovers from any error it can see. Not from an empty map.

Can frontier models write the spec? Mostly. And the failures are silent.

"getFillColor": {
  "@@function": "interpolateColor",   ← does not exist
  "value": "@@=properties.population", ...
}

Make @deck.gl/json v2 the agent interface

How does deck.gl compare with similar projects?

The contributions are already AI-assisted

Median length of a pull request description in visgl/deck.gl, bots excluded: about 400 characters through mid-2025, 1,200 to 2,500 since late 2025. Four of the six most active contributors quadrupled theirs within two quarters; two did not change. The repo's own AGENTS.md now tells agents to prepare the PR description.

The Foundation has a policy. Our repos don't say so.

But let's be real: everybody is assisted. The question is our posture.

• Naming the tool is free advertising. Assume every PR is AI-assisted and stop caring which vendor. Care about what the author verified.
• The flood is coming. PRs opened for social validation, and PRs from autonomous agents that fix bugs and improve performance without anyone asking. CARTO already runs agents like that on its own code. Who reviews them here, and who pays for that time?
• Automation is not optional. A security issue in a library like this will be found by agents first. The only defence that keeps up is agent-driven too.
• So the real design question: how do agent contributions plug into deck.gl? Verified how, tested how, budgeted how, reviewed by whom.

Disclosure is the easy part. The posture is the discussion.

Sources

• JSON Layers RFC (2018): github.com/visgl/deck.gl/blob/master/dev-docs/RFCs/v6.1/json-layers-rfc.md
• @deck.gl/json v2 tracker: deck.gl-community #596 · step 1 PR #597 · AI-forward RFCs #594, #595
• deck.gl 9.4 release notes and @deck.gl/maplibre: deck.gl/docs/whats-new
• MapLibre agent skills: github.com/maplibre/maplibre-agent-skills · Mapbox MCP servers: docs.mapbox.com/api/guides/mcp-server
• MCP Apps specification: github.com/modelcontextprotocol/ext-apps
• SQLRooms deck package: sqlrooms.org/api/deck · CARTO agentic deck.gl: github.com/CartoDB/carto-agentic-deckgl
• Open-source AI contribution policies: github.com/melissawm/open-source-ai-contribution-policies · RedMonk survey (Feb 2026)
• Experiments, raw outputs and harnesses: github.com/jatorre/deckgl-ai-ready

deck.gl on MapLibre, ten models

GLM 5.3, the open-weight model

• Its endpoint makes reasoning mandatory. Default: 24,000 tokens of thinking, no HTML, about 400 s per task. Only an explicit 8,000-token reasoning cap produced files, in 10 to 33 s.
• One earlier map pinned deck.gl@8.10.1, which does not exist.
• Choropleth: const deck = new deck.Deck(...) shadows the library global; loads Mapbox GL JS, which demands a token. Blank page.
• Hexagons: hand-rolled TileLayer reading props.bbox instead of props.tile.bbox; hexagons on black.
• Trips: mutates layer.props.currentTime and calls internal setState instead of recreating the layer; the animation dies on frame one.

PR description length by author, quarterly medians

Six most active authors, anonymized, plus everyone else (80 authors). Four of six rise from a few hundred characters to about 2,000 between 2025 Q4 and 2026 Q2; two stay flat. Bots excluded.

Adjacent talks

• 10:00 · Boyandin · SQLRooms (deck.gl JSON specs with dataset binding)
• 14:15 · Krebs & Gervang · noodles.gl, a visual programming language for deck.gl
• 14:45 · Elsinga · MLT in the deck.gl ecosystem (MapLibre)
• 16:00 · Koschinsky & Li · Skill-based design for AI spatial analysis tools
• Day 2 · kepler.gl update · Future of Vis.gl discussion 15:30