Agentic Workflow Compiler

Winner — Best Use of Cursor & Overall · tokens& Multiagents Hackathon (London, June 2026)

Paste live URLs into a node graph. Agents discover data shapes, compile tested TypeScript extractors, and freeze reusable workflows — agentic at build time, deterministic at run time.

Devpost	devpost.com/software/asdfsad — demo video, screenshots, and submission story
GitHub	github.com/ghcpuman902/agentic-workflow-compiler

Same walkthrough as on Devpost. Full demo (MP4).

---

What it is

Most AI workflow tools keep an LLM in the loop on every run: trigger → reason → tool → reason → result. That works for one-offs, but it is expensive, slow, and non-deterministic for repeated jobs like scraping the same site every day.

Agentic Workflow Compiler flips the model: agents design the automation once; frozen TypeScript runs it forever.

A Blender-inspired node editor where you:

1. Paste public URLs (Cannes Film Festival pages ship as the default demo).
2. Press Play on Discover — agent-browser inspects live pages (accessibility tree, JSON-LD, repeated DOM patterns) and suggests semantic output types (Document vs Collection).
3. Autonomous build loop — the system generates a typed extractor, tests it against live URLs, and repairs failures without manual intervention.
4. Freeze a Spider node — a reusable, deterministic workflow artefact you can re-run without calling an LLM.

Two output paths ship in this MVP:

• Document — Tavily Extract turns pages into clean Markdown (article-like structure).
• Collection — Gemini (or a Cursor agent via @cursor/sdk) writes a plain JS run(input) extractor, tested and repaired across your URL set.

---

Hackathon snapshot

This repository is preserved as the hackathon submission snapshot — the state of the project at the end of the tokens& Multiagents Hackathon, including what we demoed on Devpost.

We plan to continue development in a separate repository with a clearer name and a longer-term roadmap (scheduling, cloud workers, marketplace, etc.). When that repo is ready, this README will link to it. Until then, clone and run from here.

---

Quick start

Requirements: Node.js 20+, pnpm, Chrome for deep page inspection (via agent-browser).

git clone https://github.com/ghcpuman902/agentic-workflow-compiler.git
cd agentic-workflow-compiler
pnpm install
pnpm exec agent-browser install   # one-time Chrome download for deep inspect
cp .env.example .env.local        # then fill in API keys (see below)
pnpm dev                          # http://localhost:3121

Environment variables

Copy .env.example to .env.local:

Variable	Purpose
TAVILY_API_KEY	Document path — Tavily Extract for live web → Markdown
GOOGLE_GENERATIVE_AI_API_KEY	Schema inference, codegen, repair fallback
CURSOR_API_KEY	Stage 2 build/repair via Cursor Agent SDK (optional; falls back to Gemini)
CLICKHOUSE_URL, CLICKHOUSE_KEYID, CLICKHOUSE_KEYSECRET	Optional trace sink (local JSONL always works)
DISCOVERY_CONFIDENCE_THRESHOLD	Quick discover stop threshold (default 0.55)

Demo flow

1. Open the canvas at http://localhost:3121.
2. Connect a Text node with 2–5 public URLs to a Discover Factory node.
3. Press Play — watch Stage 1 discovery, pick Document or Collection, then let the build loop compile and test the extractor.
4. Materialize the Spider node and re-run across URLs — no LLM calls in the run phase.

See docs/03-hackathon-scope-and-demo.md for the full demo script.

---

Architecture

Paste URLs → Discover (agent-browser) → Choose output family
    → Build (Cursor agent / Gemini) → Test → Repair → Freeze Spider
    → Run (deterministic, no LLM)

Stage	What happens
Stage 1 — Discover	Quick HTTP read + optional Chrome deep inspect; infer page roles and suggest Document vs Collection
Stage 2 — Build	Generate extractor, run golden tests, auto-repair on failure
Run	Execute frozen tool across all URLs in a sandbox — zero model calls

Canonical spec: docs/06-two-stage-pipeline.md

---

Built with

Next.js 16 · React 19 · TypeScript · Tailwind CSS v4 · shadcn/ui · React Flow · agent-browser · Tavily · Google Gemini (AI SDK) · Cursor Agent SDK · ClickHouse (optional traces) · pnpm

Sponsor integrations (hackathon)

• Tavily — Document path extraction
• Cursor — Primary Stage 2 build/repair loop when CURSOR_API_KEY is set

---

Documentation

Doc	Description
docs/01-overall-context.md	Vision, thesis, roadmap
docs/02-architecture-and-models.md	System design
docs/03-hackathon-scope-and-demo.md	MVP scope and demo script
docs/06-two-stage-pipeline.md	Discover → Build pipeline spec
docs/hackathon-submission-draft.md	Original Devpost submission copy

---

What's not in this MVP

Media extraction, authenticated browser sessions, cloud scheduling, workflow marketplace, and full ClickHouse analytics UI are documented as post-hackathon roadmap items — visible in the UI but disabled or scaffolded only.

---

License

No license file is included in this hackathon snapshot. Contact the repository owner before reuse.