NemoClaw + dcode: Building Auditable AI Coding Agents in a…

Demo shows how LangChain's dcode runs inside a NemoClaw‑managed sandbox, using Nemotron 3 Ultra for inference, to fix failing tests while preserving full auditability.

Ever wondered how AI can safely code for you without stepping on regulatory landmines? This demo shows a governed sandbox that lets agents fix broken tests while staying auditable.

⚡ TL;DR: dcode runs inside a NemoClaw‑managed OpenShell sandbox, uses Nemotron 3 Ultra via Baseten, and lets developers fix failing tests with full policy and log visibility, all without leaving their normal workflow.

The Simple Version

Think of the sandbox as a secure playground where an AI agent can tinker with code. It follows strict rules set by your team, records every move, and never forces you to change how you normally code.

How It Actually Works

Installing NemoClaw

First you add the NemoClaw package to your project. This creates a controlled environment that can enforce safety policies.

🎯 Think of it this way: Installing NemoClaw is like putting a fence around a playground so only approved toys can be used.

Configuring the inference provider

Next you point the sandbox to an inference service — here, Nemotron 3 Ultra served through Baseten. This gives the agent the brainpower it needs while keeping compute resources bounded.

Provisioning the OpenShell sandbox

The sandbox spins up an isolated shell where the agent runs. Resources like CPU and memory are assigned based on a profile you select.

Connecting to the sandbox

Developers connect via a simple CLI or UI, making it feel like they are working in their regular environment.

Technical Deep Dive

Policy tier and resource profile

Each sandbox is assigned a policy tier that defines allowed actions, such as which APIs can be called or what files can be edited. Resource profiles cap compute, preventing runaway costs.

Handing a failing test to dcode

You present a failing unit test to the agent. The agent analyzes the error, writes a fix, and commits it back to the repository — all inside the sandbox.

Inspecting policy and logs

After the fix, you can review the sandbox’s audit logs to see every command the agent executed, ensuring compliance without manual code review.

Why This Matters

Teams get the productivity boost of AI coding assistants while maintaining governance, security, and traceability. Auditable logs mean no surprise breakages in production, and sandboxed resources protect budgets.

Key Takeaways

NemoClaw creates a governed OpenShell sandbox for AI agents.
Inference is handled by Nemotron 3 Ultra via Baseten.
Developers can hand off failing tests to dcode with no workflow disruption.
Full policy and log visibility provides built‑in auditability.
Resource profiles keep compute costs predictable.
The approach blends AI productivity with enterprise‑grade safety.

NemoClaw + dcode: Building Auditable AI Coding Agents in a Governed Sandbox