Microsoft’s Flint makes an argument I keep running into on the agent side: the problem with LLM-generated charts isn’t the model, it’s the target language. Vega and friends are expressive, but they force the agent to reason about scales, axes, spacing, and layout — low-level knobs that inflate token count and drag down reliability. Flint raises the abstraction so the agent describes intent and a deterministic renderer owns the composition.

• 🎯 The real pattern here isn’t charts — it’s handing agents a high-level IR and letting a compiler own the pixels
• ⚡ Verbose specs don’t fail because models can’t read them; they fail because more surface area means more places to be subtly wrong
• 🔍 Spatial composition is where LLMs are genuinely weak, so offloading it to a deterministic layer is the right seam to cut on
• ⚠️ The Vega question in the HN thread is fair: a brand-new DSL starts with zero presence in training data, and that adoption tax is real
• 💡 The same play generalizes past viz — SQL builders, tool-call schemas, workflow graphs: constrain the agent to intent, compile the messy details

The durable idea is architectural, not visual: pair a probabilistic generator with a deterministic layer that refuses to render nonsense. That’s how you get reliability out of a component that can’t guarantee it on its own — the pattern I lean on most when wiring tool use into production agents.

Will the winning agent stacks standardize on intent-level IRs like this, or will frontier models just get good enough to emit the low-level specs directly?