Swiftrender is a research prototype exploring what a single AI agent, directed by one human, can build. It is not a browser. It renders a deliberately tiny subset of HTML/CSS, has no JavaScript, and makes no speed claims. "Swift" is an aspiration, not a benchmark.
2026 produced two instructive data points for AI-led software engineering:
- FastRender (Cursor) showed that hundreds of agents can produce a browser-shaped artifact — and that a flat multi-agent structure collapses into lock contention and task avoidance unless it is replaced by a strict Planner-Worker-Judge hierarchy.
- one-agent-one-browser showed that one strong agent under tight human direction can produce a small, legible renderer in days.
Swiftrender is a deliberate bet on the second model. The repo is the
artifact; the orchestration log (PROMPTS.md, NOTES.md) is the
research output.
- Parses a small HTML subset (see
SPEC.md§2) - Cascades tag and class selectors (see
SPEC.md§3) - Block and inline layout, no floats/flex/grid (see
SPEC.md§6) - Paints to PPM, PNG (hand-rolled stored-deflate), or SVG (§8)
- 63 numbered tests, each citing its SPEC clause (
TESTS.md)
JavaScript (permanently), network, external fonts, percentages, em,
selectors beyond tag/.class, margin collapsing, baseline alignment,
anti-aliasing, image decoding, tables, forms, bidi. See ROADMAP.md.
cargo run -- fixtures/hello.html --emit png --out hello.png
cargo testDual-licensed under MIT or Apache-2.0, at your option.