DeepSeek Harness Open‑Source: Inside the V4 Pro Agent Platform

DeepSeek recently launched the V4 Pro model and, within hours, made the DeepSeek Harness source code publicly available under the MIT license. The official benchmarks showed notable gains for agents such as Terminal Bench 2.1, NL2Repo, and DeepSWE, and the code‑agent evaluation used the Harness minimal mode.

Harness serves as the runtime that bridges the model’s reasoning ability with concrete actions on files, terminals, tools, and task workflows. It can be started instantly with the command npx @deepseek-ai/dsh web, which launches a local Web UI at http://127.0.0.1:3080 after selecting a workspace and configuring a model.

The platform already covers the core skeleton of an AI‑coding tool:

Operate projects : read and edit files, run commands, use a persistent terminal and LSP.

Manage long‑running tasks : maintain plans and goals, compress context, handle background jobs.

Orchestrate multiple agents : create sub‑agents, send messages, pause or resume tasks, and integrate Claude Code or Codex as sub‑agent providers.

Connect external capabilities : support Skills, MCP, web search, custom tools, and workflows.

Control risk : provide permission presets, operation approvals, filesystem policies, and a local sandbox.

Rather than a thin wrapper around the DeepSeek API, Harness is a modular agent runtime built on the principle that “everything is a plugin.” Model adapters, tool registries, session logs, and even the agent loop itself are implemented as plugins, while the underlying Cordis layer composes services, events, and reversible side effects. The running instance, referred to as dsh, is essentially a plugin tree assembled from configuration.

This design lets developers swap individual capabilities without touching a monolithic core. The filesystem can be replaced by a local implementation or a remote sandbox; the model can be DeepSeek, Anthropic, OpenAI, or a custom endpoint; sub‑agents may run in‑process or be delegated to other products.

Session data is stored as an event log, capturing user messages, model outputs, tool calls, results, and task rounds. This persistent log enables state reconstruction, replay, and the creation of sub‑tasks, which is crucial for recovering from failures and for post‑mortem analysis.

Beyond the Web UI, DeepSeek provides a Python SDK that can launch the same runtime programmatically, specifying a workspace, session directory, and model. This makes it far more practical for teams that want to embed the agent into automated pipelines rather than clicking through a browser.

However, the repository is marked as a Developer Preview . The npm package is currently version 0.1.0-rc.6, and the GitHub repository has not yet issued an official release. The SDK example defaults to a danger-full-access mode, and the documentation warns that it should only be run in disposable code copies or containers. Although the Web UI includes sandboxing and approval mechanisms, developers must still configure permissions carefully to avoid granting excessive system access when swapping models for agents.

Consequently, the author judges that Harness is now suitable for research, trial, and plugin development, but not yet ready as a stable production product.

The open‑source release reveals a shift in the AI model landscape: competition is moving from raw benchmark scores to complete agent runtimes that handle sessions, tools, permissions, execution, collaboration, and extensibility. Harness exposes the previously hidden layer that sits beneath model inference, indicating where future product differentiation will occur.