Prompts describe work. Loops repeat work. Graphs define how work branches, verifies, remembers, and converges.

Graph Engineering is a vendor-neutral graph orchestration platform for agentic systems. Native TypeScript and Python runtimes execute one versioned Graph IR and are checked against a shared conformance corpus.

Instead of paying model tokens to coordinate a linear conversation, describe work as typed nodes and data-carrying edges. The runtime fans independent jobs out, contains failures, and converges named outputs without placing the whole job in one model context. Pure barrier and router evaluators provide deterministic decisions while scheduler-level conditional routing remains an explicit v1 goal.

The project is an early alpha. The DAG compiler, ready-queue schedulers, safe project initializer, machine-readable CLI, structured failure handling, retries, timeouts, budgets, settled barrier/router decisions, safe Mermaid/DOT rendering, pattern constructors, local event/checkpoint stores, read-only MCP server, and development progress scanner are executable today. Scheduler-integrated durable recovery and the broader v1 surface remain under active development; the repository does not silently mock unfinished capabilities.

Source-only alpha: npm and PyPI packages are not published yet. Clone this repository to try the current release candidate; registry publication remains gated on trusted publishing and package-specific security review.

Validate and inspect a research diamond without an API key:

git clone https://github.com/reacher-z/GraphEngineering.git
cd GraphEngineering
corepack pnpm install --frozen-lockfile
corepack pnpm build
node packages/cli/dist/src/cli.js validate examples/quickstart/research-diamond.graph.json
node packages/cli/dist/src/cli.js plan examples/quickstart/research-diamond.graph.json
node packages/cli/dist/src/cli.js visualize examples/quickstart/research-diamond.graph.json
node packages/cli/dist/src/cli.js init /tmp/my-first-graph --dry-run

The plan exposes two independent research nodes in the same parallel layer. See the five-minute Quickstart for expected output and failure diagnostics.

The same Graph IR becomes this execution shape:

flowchart LR
    S[scope] --> C[research: code]
    S --> D[research: docs]
    C --> M[merge + report]
    D --> M

Run that graph through the native scheduler with deterministic local handlers:

node examples/quickstart/run.mjs
uv run --project python python examples/quickstart/run.py

Python is a native runtime, not a client for the TypeScript executor:

uv sync --project python --extra dev
uv run --project python pytest python/tests

• Explicit node and edge data contracts.
• Parallel, pipeline, barrier, router, verifier, and bounded-loop topologies.
• Durable checkpoints, resume, replay, and fork.
• Deterministic plumbing; models are reserved for judgment.
• Provider-neutral adapters and deny-by-default capabilities.
• Observable runs with portable events and traces.
• TypeScript/Python semantic parity through shared conformance fixtures.

node scripts/validate-fixtures.mjs
node scripts/check-doc-links.mjs
corepack pnpm build
corepack pnpm typecheck
corepack pnpm lint
corepack pnpm test
corepack pnpm test:conformance
corepack pnpm check:packages
corepack pnpm check:packed-install
corepack pnpm audit:prod
uv run --project python ruff check python/src python/tests
uv run --project python mypy python/src
uv build --project python
python3 scripts/check-python-artifacts.py

• Concepts
• Failure modes
• Security policy
• Architecture
• Roadmap
• Changelog
• Support
• Runtime semantics
• Persistence semantics
• Primitive semantics
• 21-day delivery plan

Alpha. APIs and serialized protocols may change before the first stable release. Released protocol versions will receive explicit compatibility and migration policies.

MIT