Argus

A sourced web research agent running on Cloudflare Workflows.

• Demo: argus.coey.dev
• Paper: Argus: Evidence Assembly for Scalable Deep Research Agents

argus.coey.dev is the public docs, demo, and dogfood surface for Argus. Ask a question. Argus gathers evidence, asks Jev whether that evidence answers or disproves the original question, and researches one remaining gap at a time. It returns an answered, disproved, or blocked report with links. The evidence-loop idea is inspired by the May 2026 Argus paper; the product is source-agnostic web research.

What runs

Path	Purpose
src/routes/+page.svelte	Public demo UI and evidence-board renderer
src/routes/api/research/	Starts and polls Workflow instances
src/workflow-runtime.ts	ArgusResearchWorkflow, the durable adaptive loop
src/lib/argus/	Prompts, evidence report model, source fetch, synthesis
research/	Paper notes and product reasoning receipts

Cloudflare primitives

• Workflows — one durable research run; graph gaps create later follow-up steps.
• Workers AI — facet generation, evidence extraction, final synthesis, and typed Jev sufficiency decisions through the my-ax AI Gateway.
• SvelteKit on Workers — public docs/demo surface and Workflow API.
• Live sources — the search layer gathers source receipts; the example report shows mixed public web sources.

Run locally

Run the app through Wrangler so local Workflows + remote Workers AI execute against your authenticated Cloudflare account.

bun run dev

Open http://localhost:4173, then click Start research. If you only need rapid visual CSS/Svelte iteration without the Workflow API, bun run dev:ui starts plain Vite.

Validate

bun run check
bun run build

check runs Svelte type checking. build creates the deployable Worker bundle, generates docs/code assets, and dry-runs Wrangler deploy.

Deploy

Click the Deploy to Cloudflare button above to clone and deploy a single Worker into your own account. The Worker includes:

• the SvelteKit docs/demo/API surface;
• the ArgusResearchWorkflow binding;
• the Workers AI binding Argus uses for facets, evidence extraction, and synthesis.

The repository build runs vite build, attaches the Workflow export to the generated SvelteKit Worker entrypoint, refreshes Wrangler types, and dry-runs wrangler deploy. That same single-worker configuration is verified locally by bun run preview and deploys with:

bun run build
wrangler deploy

The Jev loop

After the first evidence pass, Jev returns one typed status:

• answered — the source-backed evidence directly answers the original question;
• disproved — strong source evidence contradicts the original premise;
• continue — one named evidence gap should be researched next;
• blocked — available evidence cannot answer or disprove the question.

Argus accepts answered only when Jev's independent evidence-sufficiency probability is at least 0.75. It accepts disproved only when the premise-disproved probability is at least 0.75. A continuation must name an available gap. The Workflow permits at most three follow-up searches. A model failure fails the Workflow instead of fabricating completion.

Jev decides evidence sufficiency. Sources carry the facts. The final writing model may only use claims already present on the evidence board.

Thesis

Research is not done because enough text exists. It is done when the evidence answers the question, disproves its premise, or establishes an honest block.