AI Workflow

Turn engineering events into inspectable, human-controlled agent workflows—and ship evidence-backed changes through the tools your team already uses.

AI Workflow is free, open source, and actively being built.

AI.Workflow.Promo.mp4

What It Is

AI Workflow is a platform for building and operating engineering-agent workflows. It turns issues, messages, alerts, webhooks, and schedules into visible multi-step runs that can research, ask for clarification, wait for approval, write code, open pull requests, respond to review, and show the evidence behind every result.

Engineering organizations need the layer around the agent: event routing, repeatable workflows, human control, delivery adapters, observability, and a way to improve successful patterns across teams. AI Workflow provides that control plane without replacing the tools your teams already trust.

Why AI Workflow

Principle	Benefit
🔍 Inspectable by default	See and customize every workflow step, prompt, decision, and artifact. No black box.
🧭 Human checkpoints	Ask for missing context, approve plans, and gate delivery at explicit points.
🔌 Keep your tools	Connect existing issue trackers, chat, alerts, repositories, and coding agents through adapters.
📈 Learn as an organization	Version workflows, govern execution, compare outcomes, and promote what works across teams.

How It Works

flowchart TD
    A["Engineering events<br/>issues · messages · alerts · webhooks · schedules"] --> B["Adapters"]
    B --> C["Custom, versioned workflow<br/>branches · loops · human steps"]
    C --> D["Agent harness<br/>context · prompts · tools · policies"]
    D --> E["Isolated sandbox<br/>Claude or Codex"]
    E --> F{"Gates<br/>clarification · approval · checks"}
    F -->|"needs input, revision, or retry"| C
    F -->|"approved"| G["Delivery<br/>PR / MR · ticket updates"]
    C --> H["Observability<br/>steps · usage · artifacts · outcomes"]
    D --> H
    E --> H
    F --> H
    G --> H

Loading

Use Cases

• 🚨 Severity response: classify an alert, gather context, and route the right workflow.
• 🔬 Investigation and research: inspect code, issues, incidents, and documentation before recommending action.
• 💬 Clarify and bounce: pause when requirements are incomplete, ask structured questions, and resume with the answer.
• ✅ Plan approval: produce an implementation plan and wait for an explicit human decision.
• 🛠️ Draft PR: implement approved work in an isolated environment and open a pull request with evidence.
• 🔁 Review and fix: react to checks or review feedback, apply changes, and return the work for review.

Target Capabilities

• Jira and Linear issue triggers and lifecycle updates.
• Slack messages, alerts, webhooks, and schedules as workflow inputs.
• GitHub and GitLab delivery through pull requests and merge requests.
• Claude and Codex as interchangeable execution agents.
• Visual, typed, multi-step workflows with branching, loops, and human steps.
• Prompt and workflow versioning with history, comparison, and restoration.
• Per-run monitoring for steps, outcomes, timing, model usage, and artifacts.
• Promotion of proven workflows plus team and outcome performance views.
• Customer-controlled deployment, data, credentials, and execution infrastructure.

How AI Workflow Is Different

Legend: ✅ native/current · ◐ partial, adjacent, or tier-dependent · — not publicly documented / not the product focus.

Product	Open-source core	Customer-controlled application deployment	Visual multi-step workflow graph	Jira → PR	GitHub + GitLab delivery	Structured human clarification/approval	Per-run step and usage telemetry
AI Workflow	✅	✅	✅	✅	✅	✅	✅
Claude Tag	—	—	—	—	◐	◐	◐
Ellipsis	—	◐	◐	—	—	◐	✅
Tembo	—	✅	◐	✅	✅	◐	◐
OpenHands	✅	✅	◐	◐	✅	✅	✅
Rovo Dev	—	—	◐	✅	—	◐	◐
DX	—	◐	◐	◐	◐	✅	✅
LinearB	—	—	◐	—	◐	✅	◐
Swarmia	—	—	—	—	—	—	◐
Jellyfish	—	—	—	—	—	—	◐

Architecture and Deployment

AI Workflow separates event adapters, typed workflow definitions, durable orchestration, isolated agent execution, delivery adapters, and observability. A run keeps its inputs, step state, human decisions, outputs, usage, and delivery evidence connected from trigger to outcome.

Deploy the application, durable orchestration, workflow and run state, credentials, and isolated execution in infrastructure you control. The open-source dashboard provides workflow authoring, run inspection, usage visibility, and operational controls.

See workflow definitions for the graph model and block catalog.

Get Started

Follow SETUP.md for local development, environment variables, and deployment.

Version-control setup:

• GitHub App setup
• GitLab setup

Repository Map

ai-workflow/
├── apps/
│   ├── worker/      # Events, orchestration, agents, adapters, and APIs
│   ├── dashboard/   # Workflow authoring, observability, and administration
│   └── shared/      # Shared contracts and workflow conditions
├── docs/            # Product, integration, deployment, and testing guides
├── SETUP.md
└── package.json

Contributing

This repository is a pnpm workspace. From the repository root:

pnpm install
pnpm dev
pnpm dev:dashboard
pnpm typecheck
pnpm test
pnpm build

Roadmap

Work in progress.

• Broader adapter coverage for Linear, Slack, alerts, webhooks, and schedules.
• Governance for audit history, policy enforcement, and execution budgets.
• Native prompt lifecycle management.
• Organization-wide promotion, canary rollout, and rollback for workflows.
• Team and outcome analytics.
• Portable deployment beyond Vercel.

Priorities may change as the project develops.

License

AI Workflow is free and open source under the MIT License.