Replace the agent loop when the default gets in your way
DeepSeek Harness lets you replace the part of an agent that decides what happens next. Its Cordis architecture makes the model adapter, tool registry, session log, and agent loop plugins. That matters if your team needs a different approval policy or execution environment without maintaining a fork of the whole application. An ordinary developer who just wants help editing a repository has less reason to learn this machinery.
The source checkout contained 14,899 files and roughly 1,214,194 lines in our lab's source scan. Those figures put some weight behind the word "harness": adopting the development checkout means entering a substantial TypeScript monorepo. The architecture reference explains how plugins register services and events, then undo their registrations when unloaded. Read that before deciding to write your own extension. The attraction is control over behavior, with a sizeable codebase attached.
What happened when we ran it
Our run installed 1,397 pnpm packages in 36 seconds, used 2,065 MB on disk, and built successfully in 350 seconds. We cloned commit d743267 on October 13, 2026 into a fresh Debian container with 3 CPUs and 8 GB of RAM, using lab-node:22. The container was unprivileged and held no secrets. Installation and compilation both completed, which supports trying the source route on a comparable machine.
The tests timed out at 900 seconds. The supplied tail still shows passing checks for session search, filesystem integration, pending submissions, and YAML reload behavior. Other passing checks concern preserving a user's pre-existing uncommitted work and recording changes after tools settle. The tail contains no final suite summary or exception identifying the reason for the timeout. We cannot call the whole suite green, invent a total passed count, or blame a missing system package.
The scan found 23 CI workflow files and monorepo workspaces, with no Dockerfile or top-level tests directory. That last signal needs care: the log explicitly names test files inside package directories. Tests exist, and some ran successfully. The useful finding is the incomplete full run. For a team contributing changes, budget time to identify the relevant checks and establish a repeatable local result before relying on the whole suite as a release condition.
A running browser still needs a model and workspace
The 350-second source build prepares artifacts that pnpm dsh web reuses. The README also offers npx @deepseek-ai/dsh web, which launches a browser interface on loopback. We measured the checkout path, so the npm shortcut carries no timing claim here. Once the browser opens, the Web UI guide tells you to save a DeepSeek API key and choose a workspace. The composer stays unavailable until that directory is selected.
Other providers and custom model endpoints are supported through the Models settings. A custom endpoint needs a provider identifier, base URL, protocol, credential, and model entry. The provider guide warns that a correct key and reachable gateway can still produce rejected requests when their request formats differ. Some compatibility settings require editing cordis.patch.yml. Keys live in the Harness home's credentials file, and OAuth providers are unsupported in this settings flow.
Source development carries a large rebuild cost
A 126.3 MB checkout became a 2,065 MB installed environment in our sandbox. Keep that footprint in mind before choosing the source tree for a disposable machine or shared development image. The development guide also requires Node's TypeScript stripping support and a Corepack-enabled pnpm setup. Disabling that runtime capability can prevent the build scripts from loading their configuration. The README's short command sequence leaves these details to the contributor documentation.
Windows and WSL development have their own instructions: keep the checkout, dependencies, and toolchain in the same operating system environment. The guide warns about cross-filesystem overhead and platform-specific binaries. Our 350-second Debian build establishes nothing about a Windows workstation's experience. If Windows is your deployment target, validate that path separately rather than treating a Linux build as portable proof. The documentation earns credit for spelling out this boundary.
Developer preview excludes a production security promise
The 900-second timeout is one reason to qualify adoption, and the project's own warning is another. Its safety notice says the software has not undergone a security audit and must not be treated as secure or production-ready. Model-generated commands and third-party plugins can reach whatever files, credentials, processes, and network access you provide. Permissions and approval prompts reduce exposure without guaranteeing isolation. Choose a disposable environment and narrowly scoped access for an evaluation.
The README also promises compatibility-breaking changes during developer preview. That rules out a comfortable recommendation for teams demanding stable extension contracts. GitHub recorded the last push on October 9, 2026 and listed zero open issues and PRs, while the latest-release endpoint returned no release. Feedback goes through Discussions, so the empty issue list provides no evidence that users have no problems. The recent push supports ongoing work, though issue activity cannot establish support responsiveness here.
Pick a coding assistant unless you need to change its internals
DeepSeek Harness's 1,397-package source setup makes most sense when replacing adapters or execution behavior is part of your job. Aider is a better starting point for terminal pair programming with Git-managed edits. OpenCode suits someone who wants a coding agent with separate build and read-only planning modes. Neither comparison carries a speed claim. Try Harness when you can name the plugin boundary you need to change and accept maintaining that choice through breaking updates.

