DeepSeek Harness field guide
From first run to a working mental model
Ten verified guides across setup, architecture, and real-world use.
One path from evaluation to integration
- 01
Getting Started
Build the right mental model, then complete your first run.
3 articles - 02
Architecture
Understand plugins, composition, session events, and capability seams.
4 articles - 03
Usage
Bring model configuration, the CLI, and the Python SDK into real workflows.
3 articles
A composable plugin tree, not a fixed kernel
Profiles assemble Bundles and Patches. Cordis plugins provide session, model, and tool capabilities. The Agent Loop advances model requests and tool results through Session Events.
- Profileweb / headless
- Plugin treeService · Event · Effect
- Agent Loopmodel ↔ tools
Getting Started
Build the right mental model, then complete your first run.
- 01
What Is DeepSeek Harness? From Agent Runtime to ‘Everything Is a Plugin’
Separate the harness from the model and API, inspect its shipped compositions, and run a bounded evaluation before considering production use.
14 min read - 02
Launch the DeepSeek Harness Web UI in Five Minutes
Start the browser interface in a disposable workspace, configure a model safely, run an observable first task, and diagnose failures by layer.
14 min read - 03
Run from Source: Clone, Build with pnpm, and Start Developing
Build a reproducible contributor checkout, understand generated artifacts, launch the Web and headless profiles, and diagnose source failures without hiding them.
16 min read
Architecture
Understand plugins, composition, session events, and capability seams.
- 04
Everything Is a Plugin: Cordis, Context, and Reversible Effects
Follow a capability from Context and Fiber through Service, Event, Effect, configuration, mounting, and verified disposal.
17 min read - 05
Profiles, Bundles, and Patches: How the Harness Is Assembled
Inspect the complete composition, install profile-owned extensions, apply whole-row patches safely, and prove precedence before boot.
17 min read - 06
Agent Loop and Session Events: How One Turn Runs
Trace queued input through turns, steps, model streaming, ordered tool results, durable replay, SDK observation, cancellation, and recovery.
18 min read - 07
Capability Seams and the Tool Execution Pipeline
Separate Definition, Provider, and Consumer roles, then trace policy, approval, guards, dispatch, sandboxing, results, and concurrency.
18 min read
Usage
Bring model configuration, the CLI, and the Python SDK into real workflows.
- 08
Configure DeepSeek and OpenAI-Compatible Model Providers
Configure native, catalog, and custom routes while keeping credentials write-only, model metadata honest, and sessions reproducible.
18 min read - 09
Headless CLI Automation: Bring the Harness into Scripts and CI
Run one fresh persisted task from a controlled workspace, interpret stdout and exits, inspect composition, enforce permissions, and shut down safely.
18 min read - 10
Python SDK in Practice: Connect a Client and Run Session Tasks
Install the paired runtime, configure workspaces and sessions, interpret activity intervals and notifications, and close safely.
18 min read
Open-source companion
Keep the field guide beside your code
The multilingual handbook adds recipes, troubleshooting paths, architecture notes, and verified upstream updates.Sources and boundaries
SandBase independently prepared this guide from the official DeepSeek Harness open-source repository. It is not official DeepSeek content and does not imply a partnership. Commands and interfaces reflect only the developer-preview state on the verification date.
Visit deepseek-ai/deepseek-harness ↗