AgentKit Studio Documentation
Everything you need to build, run, deploy, and observe enterprise agents on DeepSeek Harness.
Hand the idea to an agent
AgentKit Studio is a product plugin for DeepSeek Harness (DSH): author agents, bind enterprise resources, run native conversations, and deploy agents to the cloud from one workbench. This documentation covers everything from the first minute to production operations.
Start here
5-minute quickstart
Install the desktop app, configure a model, create your first agent, and run a real conversation.
Core concepts
Agent, revision, build, run, session, resource binding — the six words that explain the product.
Guides
Task-oriented playbooks: bind MCP servers, deploy to the cloud, configure approvals, wire telemetry.
Reference
CLI, configuration, environment variables, error codes, template schemas — generated from code.
Documentation map
| Section | What you will find | Who it is for |
|---|---|---|
| Getting started | Install, quickstart, first conversation | Anyone opening the product for the first time |
| Concepts | Product model, three runtimes, resources, glossary | Anyone deciding which capability to use |
| Guides | Step-by-step playbooks with verification | Developers and operators who use it daily |
| Reference | Contracts, enums, error codes, commands | Integrators, plugin authors, on-call |
| Runtimes and deployment | Runtime comparison, managed runtime lifecycle, degradation | Owners of shipping and reliability |
| Ecosystem and plugins | DSH plugin authoring, plugin market, expert templates | Anyone extending the product |
| Releases and migration | Release notes, compatibility, upgrade steps | Anyone tracking versions |
| Help | FAQ, troubleshooting, known limitations | Anyone who is stuck |
| Contributing | Dev environment, gates, patch governance, writing style | Contributors |
What the product does
Agent authoring
Three runtimes
Resource bindings
Cloud deployment
Experts and teams
Plugin ecosystem
Docs ship with the product
This site publishes automatically from the main branch. The "last updated" stamp comes from Git,
and prose is reviewed in the same pull request as the code — there is no window where the docs
describe a product that no longer exists.