Enterprises are running into the same wall with AI coding agents: the tools that write code fastest are usually the ones IT trusts least with sensitive codebases. Legacy modernization projects — COBOL migrations, .NET upgrades, decade-old frameworks nobody wants to touch by hand — are exactly where agentic coding tools could help most, and exactly where a black-box agent with no audit trail is a hard sell to a CISO.

That tension is why LangChain’s dcode, an open-source terminal coding agent, is worth a second look this month, even though it isn’t new. Deep Agents Code, its full name, traces back to Deep Agents CLI, which LangChain introduced in October 2025 as a framework for building AI agents with persistent memory. The dedicated coding agent built on that framework, dcode, first shipped on PyPI at the end of April 2026. It has been updated constantly since — more than 55 releases, with the latest landing just this week. So no, it isn’t brand new. What is new is the attention it’s getting, thanks to a recent partnership that puts it inside a governed sandbox for enterprise code.

dcode runs from the terminal and works like Claude Code or Cursor’s agent mode, but it’s built to be model-agnostic. Teams can point it at any large language model that supports tool calling and switch providers without rebuilding their setup. It maintains persistent memory across sessions, supports customizable skills that shape its approach to a task, and can delegate work to subagents for parallel execution. Approval gates allow a human to sign off before the agent executes shell commands or touches files, and they can pull in external tools via Model Context Protocol servers. LangSmith handles tracing for teams that want visibility into what the agent actually did and why.

That last point — visibility — is the piece that’s driving renewed interest. In July, LangChain and NVIDIA released a NemoClaw blueprint that pairs dcode with NVIDIA’s Nemotron 3 Ultra model inside a sandboxed, governed environment built for sensitive codebases. The setup uses deny-by-default networking, per-request approval for any outbound connection, full audit trails, and per-session snapshots, with credentials kept entirely outside the sandbox. LangChain describes the goal plainly: give teams the capability of an agentic coding tool “without the risk, the lock-in, or the data exposure.” The primary use case is legacy modernization — the COBOL-to-Java and framework-upgrade work that’s been sitting on backlogs for years because nobody wanted to hand it to a tool they couldn’t audit.

Mitch Ashley, an analyst at Futurum Group, says that record is the whole ballgame. “Platform teams don’t block coding agents over code quality,” Ashley said. “They block an agent with shell access to production-adjacent systems that lacks a log of its changes. This blueprint gives a platform lead the record a change advisory board asks for.”

For DevOps and platform teams, that governance model is the more interesting story than the agent itself. Coding agents have advanced quickly over the past two years, but most enterprise holdouts no longer doubt that the models can write decent code. dcode’s approval gates and the NemoClaw sandbox are a direct answer to the concern that actually stalls adoption, and they matter more to a platform engineer deciding whether to greenlight a pilot than any benchmark score.

The project isn’t finished making that case, either. An open roadmap discussion on GitHub lays out where dcode still falls short of production-grade platform tooling: no first-party Kubernetes operator yet for multi-tenant, autoscaled deployments, no Language Server Protocol integration for automatic error detection and self-correction, and thinner role-based access controls than some competitors offer. Today, dcode runs mainly as a CLI and terminal UI backed by SQLite — solid for a single developer, less so for a platform team trying to run it as a shared service across dozens of engineers. The same roadmap thread stacks dcode up against OpenCode and other open coding agents on exactly those platform dimensions, which is a useful reminder that “open source” and “enterprise-ready” aren’t the same claim.

Ashley also flags where the governance actually lives. “NVIDIA’s runtime holds the governance, not the open harness,” he said. “That makes the audit trail a vendor commitment on a 12 to 24 month horizon. Test whether the logs survive a runtime swap before you standardize.” It’s a fair caution for a platform team evaluating the blueprint: dcode itself is MIT-licensed and portable, but the audit trail and sandbox controls that make NemoClaw compelling are part of NVIDIA’s runtime. Those gaps are being actively worked on the dcode side, which tracks with a project that’s shipped a new release almost every week since spring — but the governance layer is worth testing independently before anyone standardizes on it.

None of that makes dcode a household name yet. It’s an actively evolving open-source project, not a fresh launch, and its adoption numbers are still modest next to the big commercial coding assistants. But the pairing with NemoClaw is a signal worth watching. It suggests that the next phase of the coding-agent conversation won’t be about which model writes the cleanest function. It’s going to be about which agents come with the guardrails that enterprises already require in near-production code — approval workflows, audit trails, a sandbox that keeps credentials out of reach — and how much of that governance a team actually owns versus rents.