Harness Handbook

Independent, source-backed, agent-first

Operate the harness, not just the model.

Understand DeepSeek Harness from first run to production boundaries.

Star the handbook Download field guide
DeepSeek Harness Handbook map showing agent runtime, tools, policy, sessions, MCP, and production paths

Five-minute first run

Start with evidence.

Launch the official Web profile from a disposable workspace, choose a model, and begin with a bounded read-only task.

Terminal
npx @deepseek-ai/dsh web
  1. Choose the workspace.Select the exact disposable repository the agent may inspect.
  2. Configure the model.Use a limited credential with an environment spending cap.
  3. Ask for evidence.Require real file paths and stop before writes or external effects.
  4. Inspect the graph.Run dsh --profile web --dump-config before changing policy.

A runtime is a composition.

A profile resolves bundles and patches into one Cordis graph. The model is one adapter inside that graph, not the product boundary.

Model
Provider, endpoint, credential reference, and reasoning route.
Effects
Tools, approval policy, sandbox, and independent verification.
Session
Durable events that reconstruct model-visible history.
Surface
Web, headless, SDK, or a custom client over the same runtime.

Pick the path that matches the job.

Each guide states the expected evidence, failure branch, and safety boundary.

CurrentCheck the rc.6 field statusKnown boundaries, evidence, and safer next actions InteractiveRoute a failure in 30 secondsPick the first broken boundary and collect the right evidence Incident noteFix an empty remote Web UIDiagnose secure-context and crypto.randomUUID failures Incident notePrevent subagent route driftWhy a child agent may use the wrong provider and credential Web UIExplore interactivelyWorkspace selection and a first safe task HeadlessAutomate one bounded jobExit codes, artifacts, diffs, and CI evidence Python SDKEmbed the runtimeJSON-RPC lifecycle and platform boundaries MCPConnect external capabilitiesProcess, transport, handshake, registration, policy TroubleshootingDiagnose by symptomSeparate provider, loop, policy, sandbox, and tool failures

A prompt is an instruction. The boundary lives in policy, isolation, and evidence.

Pin the runtime. Review the resolved graph. Separate credentials by environment. Test denial, timeout, partial completion, and cleanup.

Open the operator cheat sheet

Build on facts that can be replayed.

The handbook is independent, open source, and verified against primary DeepSeek Harness sources.

Read and star on GitHub