DeepSeek has open-sourced Harness, an agent runtime that treats almost every part of an AI agent, including the reasoning loop itself, as a replaceable plug-in. That makes it a significant piece of computing infrastructure, but not a ready-made coding assistant for tomorrow morning.
DeepSeek Watch analysis
What happened
KDnuggets reports that Harness is available as the dsh command-line tool and can be installed through npm. Its profiles support web, headless, terminal and rescue modes, while plug-ins can be added through the same package mechanism rather than through a separate extension system.
The most concrete detail is the architecture. A test of the project’s –dump-default-config option showed 152 separately named plug-ins in the default web profile, covering elements such as the interface, approvals, sub-agents, scheduling and model configuration. The article also reports a clean missing-credentials error when the headless mode was run without an API key, rather than an impressive but useless stack trace.
The project is described as a developer preview. KDnuggets says it supports roughly 40 model providers, uses operating-system isolation through tools including Bubblewrap, Landlock and Apple’s Seatbelt, and keeps append-only session logs. Those are reported project characteristics, not an independent security audit.
Why it matters
Most agent tools let developers adjust the model, tools or prompts while leaving the central loop buried inside the application. Harness makes that loop another package that can be swapped through configuration. In theory, developers could change how an agent plans, delegates or recovers without forking an entire application.
That is a meaningful shift in where the contest may happen. The model is still important, obviously, but the layer deciding how tools, memory, permissions and sub-agents fit together could become the more durable piece of infrastructure. The plumbing has acquired ambitions, which is usually when plumbing gets expensive.
Our read
This is worth the attention of people building agent platforms, sandboxes or developer tooling. It is not yet a sensible recommendation for anyone seeking a polished replacement for Claude Code or Codex. The architecture is the story, while the ecosystem, compatibility and security record still need to prove themselves.
What to watch
- Whether third-party developers can replace the agent loop without fighting undocumented interfaces.
- How the sandbox behaves under hostile or deliberately awkward tool calls.
- Whether the developer-preview API stabilises before the plug-in ecosystem fragments.
- Whether model-agnostic support remains practical as providers change their interfaces.
Discussion spark: Does making the agent loop a plug-in create a genuinely more open platform, or simply move the lock-in to whichever plug-ins become unavoidable?
Sources and evidence
- What I’ve Learned About DeepSeek Harness – KDnuggets (24 September 2026, 12:09 UTC)
not affiliated with or endorsed by DeepSeek