The Human-Controlled AI for Developers

Two cores, one workbench: a visual Workflow builder and a real agent Harness — right inside VS Code.

_{AI agent harness · agent workflow builder · human-in-the-loop · context engineering · Claude Code · Cursor · Copilot · Aider}

Not just another fancy agents tool. NWA Agent is a human-controlled AI that gives developers full control over how multiple agents think, work, adapt, and deliver — in VS code. It is built to amplify developers, not replace them.

---

🧩 Workflow is not enough. Meet the Harness.

Everyone is shipping "workflows." A workflow is the plan — it says what the agents do and in what order. That's necessary. But run on its own, a workflow is a black box: it starts, it runs, and you hope for the best.

The thing that actually decides whether an agent succeeds isn't the plan. It's the harness — the scaffolding around the model:

• 👀 what the agent sees (context delivery)
• 🛠️ what it can touch (tools, permissions, sandboxes)
• ✋ when it must stop and ask you (verification & human-in-the-loop)
• 🧠 what it remembers (project memory & state)
• 🔁 what happens when things go wrong (feedback loops & recovery)

💡 The models are converging. When everyone has access to the same frontier models, "which model?" stops being the moat. How good is your harness? becomes the real competitive advantage — the same model can go from ~60% to ~98% task success depending entirely on the harness around it.

NWA Agent gives you both, as first-class citizens:

        WHAT to do, in what order            HOW it actually runs, safely
        ┌───────────────────────┐            ┌───────────────────────────┐
        │       WORKFLOW        │     ×      │          HARNESS          │
        │  steps · parallel ·   │  ───────►  │  context · tools · memory │
        │  per-step models      │            │  human-in-the-loop · state│
        └───────────────────────┘            └───────────────────────────┘
                              \                /
                               ▼   reliable   ▼
                          AI that ships real work

---

✨ Why this repo exists

Read each row left-to-right: a real pain → the plan that addresses it → the control that makes the plan actually land → the payoff you feel.

Pain Point	The Workflow plan	The Harness control	What you actually get
🤖 Shallow "agent" tools that live outside your editor and break the moment real code is involved	A visual pipeline of concrete steps — collect → generate → test → review — modeled on how you really ship	Runs natively in VS Code against your real workspace files, terminal, and Git — no copy-paste bridge	AI that works in your repo, not in a sandbox demo — the output is the change, ready to review.
🏃 Black-box runs that sprint start-to-finish and only show you the wreckage afterward	Steps are observable units — you watch each one start, progress, and finish, with live status and token cost	Human-in-the-loop review gates halt the run and hand control back to you before the next step commits anything	Catch a wrong turn at step 2 instead of debugging the whole run at step 9 — you approve, then it continues.
🔁 No steering wheel mid-run — your only options are "let it finish" or "kill it and restart"	Steps stay editable while idle — reorder, re-scope, or swap a step's role between runs	Mid-flow intervention: pause at a gate, rewrite the brief or context, then resume from there — earlier work is kept	Change direction without throwing away the 20 minutes of context the agent already built.
🧠 Context chaos — the agent either sees nothing useful or drowns in the whole repo	Each step carries its own instructions and model, so the right brain gets the right job	Layered context delivery (auto-loaded rules · on-demand docs · hot data) + persistent project memory that survives across runs	Agents start informed instead of from zero — fewer hallucinations, less re-explaining the same thing.
🎯 One-size model wastes money on easy steps and underthinks the hard ones	Per-step model + reasoning-speed selection baked into the workflow	The harness routes each step to the chosen model/runtime and enforces it at execution	Spend deep reasoning where it matters, fast-and-cheap where it doesn't — cost and quality both go up.
🔒 Rigid, copy-pasted workflows that silently drift apart and can't be reused	Workflows are composable from @nwa/step-* packages on a shared SDK, and import/export across projects	Each step is an isolated, core-free package with a single execute entry — change one without breaking the rest	Build a workflow once, reuse and version it everywhere — your team's process becomes a real artifact.
⚙️ Speed vs. control is a false trade-off most tools force on you	Multi-agent + parallel groups run independent work concurrently	Oversight & guardrails — gates, memory, and per-step limits keep the speed from running off a cliff	Frontier-model speed with your judgment in the loop — fast and trustworthy, not one or the other.

---

🪴 Quick demo

---

📦 Installation

The final installation path and package source will be updated before release.

From VS Code Marketplace

1. Open VS Code.
2. Go to Extensions (Ctrl+Shift+X / Cmd+Shift+X).
3. Search for "New Workbench Agent".
4. Click Install.

From VSIX

Use the packaged extension path after it is published or built locally:

code --install-extension <path-to-new-workbench-agent.vsix>

From Source

npm install
npm run build

Then press F5 in VS Code to launch the Extension Development Host.

---

🚀 Quick Start

1. Open the Agent Manager

1. Open the Command Palette with Ctrl+Shift+P or Cmd+Shift+P.
2. Run "NWA: Open Agent Manager".
3. Choose the AI tool you want to configure.
4. Select the departments and agents that match your workflow.
5. Click Install Agents.

2. Use a Quick Setup Preset

1. Open the Command Palette.
2. Run "NWA: Quick Setup".
3. Select one of the available presets:
   • 🚀 Full Stack Developer
   • ⚡ Rapid Prototyper
   • 🎨 Design-First
   • 📈 Growth-Focused
   • 🏢 Enterprise Team

3. Build a Workflow + wire its Harness

1. Open the NWA Workflows view and create a workflow — add steps, parallel groups, and pick a model/speed per step.
2. Open the NWA Task Manager to run it, watch progress, and step in at human-in-the-loop review gates.
3. Use Claude Resource Manager to give each agent the right context (rules, docs, hot data).

4. Manage From the Sidebar

1. Open the NWA Activity Bar view.
2. Browse the Tools, Task, Memory, and Component Browser views.
3. Run NWA: Open Agent Manager or NWA: Skills selection when you need to install more resources.

---

📚 Claude Resource Manager

This is your context harness. Initialize Claude Code context resources by selecting files from each layer. You can select one file, several files, or every file, then use the floating Install action to create the selected resources in your workspace.

Layer 1: Auto-loaded Rules

Layer 2: On-demand Docs

Layer 3: Hot Data

SKILL.md is installed to .claude/skills/SKILL.md. If a previous SKILL.md exists there, it is renamed to SKILL.md-old before the new file is created.

---

⚙️ Configuration

Settings

Open VS Code Settings (Ctrl+, or Cmd+,) and search for "NWA":

Setting	Default	What it does
nwa.defaultTool	cursor	Default AI tool for agent configuration — cursor, claude-code, copilot, aider, or universal.
nwa.defaultFolder	""	Custom default folder name (empty = tool-specific defaults).
nwa.autoRefresh	true	Automatically refresh the agent list when files change.
nwa.showWelcome	true	Show the welcome message on first use.
nwa.defaultDepartments	["engineering","design"]	Pre-selected departments — design, engineering, marketing, product, project-management, studio-operations, testing.
nwa.favoriteAgents	[]	Starred agent IDs, managed through the UI.
nwa.taskDocumentsFolder	.project/docs	Workspace-relative folder where the Task Manager stores markdown documents (converted with markitdown).
nwa.componentPaths	src/components, src/atoms, …	Workspace-relative paths the Component Browser and Figma catalog scan for Vue / React / Angular / Svelte components.

Commands

All commands are available from the Command Palette (Ctrl+Shift+P / Cmd+Shift+P).

Agents & setup

• NWA: Open Agent Manager — open the visual agent manager.
• NWA: Quick Setup — install agents from a preset.
• NWA: Initialize Agents — run guided setup.
• NWA: View Agent Details — inspect an agent.

Context & skills (the harness)

• Install Helpers — set up the Claude Code environment.
• NWA: Skills selection — open the Claude Resource Manager to install context resources.

Workflows & tasks

• NWA: Task — open the Task Manager to run and track work.
• NWA: Workflows — open the visual workflow builder.
• NWA: Workflow Settings — configure workflow defaults.

Memory

• NWA: Add Memory — save a memory manually.
• NWA: Search Memory — search project memories.
• NWA: Show All Memories — list every stored memory.
• NWA: Clear All Memories — remove all memories.

Component Browser

• NWA: Refresh Component Browser — re-scan component paths.
• NWA: Add Component Scan Path / NWA: Remove Component Scan Path — manage scan paths.

Figma MCP bridge

• NWA Agent: Start Figma MCP Bridge / Stop Figma MCP Bridge / Show Figma MCP Status.

🐛 Troubleshooting

Agents not working with Cursor

1. Restart Cursor after installing agents.
2. Make sure files exist in .cursorrules/.
3. Use @ to mention agent files.

Claude resources not appearing

1. Open NWA: Skills selection.
2. Select the resources you want to install.
3. Click the floating Install action.
4. Refresh the VS Code Explorer if needed.

Extension not activating

1. Check your VS Code version. This extension requires 1.85.0 or newer.
2. Reload the window with Ctrl+R or Cmd+R.
3. Check the Output panel for errors.

---

🤝 Contributing

Found a bug or have a feature request?

• Critical project discussions: GitHub Discussions
• 🐛 Report bugs: GitHub Issues
• 💡 Suggest features: GitHub Discussions
• 📖 Documentation: GitHub Wiki

---

📄 License

This project is licensed under the MIT License. See the LICENSE file for details.

---

Made with love by b0yblake