Orchestrate AI agents to ship reliable software.

Newton turns AI coding agents from unpredictable assistants into dependable contributors. Instead of one-shot prompts and hoping for the best, Newton runs them inside deterministic workflows with explicit goals, checkpoints, approvals, and verifiable completion.

AI agents are powerful, but raw agent runs are hard to trust in real engineering work:

• Output drifts from the requested goal.
• There is no clear definition of "done".
• Failures are opaque and not resumable.
• Humans lose oversight of what the agent actually changed.

Newton solves this by treating agent work as a workflow graph: every step has inputs, outputs, success conditions, and recovery semantics. Agents become operators in that graph, side-by-side with shell commands, GitHub actions, and human approvals.

• Deterministic workflows authored in YAML, with linting, dry-run preview, and a visual graph.
• Agent operators for the major coding agents (Claude, Codex, Gemini, OpenCode, generic SDK agents) with uniform quota and error handling.
• Goal gates and terminal tasks that decide when a run is actually complete, not just "finished".
• Checkpoints and resume so long agent runs survive crashes, restarts, and human breaks.
• Human-in-the-loop approvals and multiple-choice decisions wired directly into the graph.
• GitHub integration for PR creation, review, and project board updates as first-class steps.
• Artifacts and logs captured per task, ready for audit, replay, or downstream tooling.
• Sub-workflows to compose large pipelines from small, reusable graphs.

• Agent-driven feature delivery. Plan, implement, test, open a PR, and chase reviews, with humans gating the risky parts.
• Batch coding tasks. Queue many specs and have Newton drive an agent through each one with consistent setup, branching, and post-run hooks.
• Release and operations runbooks. Encode deploys, migrations, and incident playbooks as graphs instead of brittle scripts.
• Evaluator / advisor loops. Iterate on a solution with measurable scoring until a goal is met, instead of fixed-iteration prompting.
• Multi-agent pipelines. Combine specialized agents (planner, implementer, reviewer) under one orchestrator.

1. Describe the work in a workflow YAML: tasks, operators, dependencies, and completion rules.
2. Run it with newton run. Newton schedules tasks, invokes agents and tools, captures artifacts, and checkpoints progress.
3. Stay in control. Use newton monitor for a live terminal UI, answer approval prompts, inspect artifacts, and resume from any checkpoint.
4. Verify completion. Goal gates and terminal tasks decide success; failures get stable error codes you can act on.

macOS / Linux:

brew tap gonewton/cli
brew install newton

Windows:

scoop bucket add gonewton https://github.com/gonewton/scoop-bucket
scoop install newton

Then bootstrap a workspace:

newton init .
newton run

• CLI reference and operator docs: gonewton/newton
• Workflow templates: gonewton/newton-templates
• Homebrew tap: gonewton/homebrew-cli
• Scoop bucket: gonewton/scoop-bucket

Newton is released under the Apache License 2.0. See each repository’s LICENSE file for the full text.

Contributions, issues, and workflow recipes are welcome.