Wujie HarnessDocs
Wujie HarnessDocs
WelcomeGetting started

Basics

Workspaces and membersIssuesChannelsProjectsLabelsWorkflowsInbox and subscriptions

Agents

Meet the agentsCreate and configure an agentAssign issues to agentsChatSquadsSkills

Automation

AutomationsRuntimes

Knowledge

DocumentsKnowledgeCapability CenterConnectors

Governance

Governance

Create and configure an agent

Two required fields — a name and a runtime — plus every optional setting worth knowing.

Creating an agent takes only two things: a name and a choice of AI coding tool. Everything else — instructions, model, environment variables, visibility, concurrency — is optional and editable at any time. Get it running first, tune it later.

Pick an AI coding tool

Each agent uses an AI coding tool exposed by a connected runtime. For a local runtime, the tool is installed on the target machine; remote and cloud runtimes provide the tool from their respective execution nodes. The most common choices are:

ToolGood for
Claude CodeAnthropic's official tool, most complete feature set; best first pick
CodexOpenAI, the mainstream alternative
CursorUsers in the Cursor editor ecosystem
CopilotTeams leveraging their GitHub account entitlements
CodeBuddyTeams using Tencent's Claude Code-compatible CLI

Additional providers are supported. The creation flow displays the options returned by the runtime you have connected; treat that live list as the source of truth instead of a hard-coded provider count.

Writing system instructions

System instructions are prepended to every task — they tell the agent what role it plays and what rules to follow:

You're a frontend code-review agent. When an issue comes in, read the diff first. Focus only on:
- Styling issues (tailwind class names, box model)
- Accessibility (a11y)
Don't change code — leave suggestions in a comment.

Leave it blank and the agent uses the native behavior of its underlying tool with no extra constraints.

Model and runtime settings

  • Model — most tools support picking a model (Claude Code offers Sonnet and Opus, for example). Leave blank for the tool's default. Changing the model only affects new tasks; already-dispatched tasks finish with the model they started on.
  • Concurrency limit — how many tasks the agent can run in parallel (default 6). Tasks that hit the cap queue up rather than get rejected. Changing it doesn't interrupt running tasks.

Custom environment variables

Custom env vars are injected at task execution time — typical for API keys or alternate endpoints:

ANTHROPIC_API_KEY = sk-...
ANTHROPIC_BASE_URL = https://my-proxy.example.com

System-critical variables (PATH, HOME, USER, SHELL, and anything starting with WUJIE_) cannot be overridden.

Environment variable keys can be shown in agent configuration, but values are readable only by the agent owner. Workspace admins do not receive an owner bypass. Even so, don't put high-value secrets here: use dedicated, limited-scope credentials for agents, and rotate them regularly.

Visibility

  • Workspace — workspace members can use it in their workflows
  • Private (default) — only the agent owner can trigger it

Private agents are still visible to every member — name, description, and visibility remain discoverable. Private restricts who can trigger execution, not who can find the agent. Agent edit, archive, and restore actions are owner-only; workspace admins do not bypass that rule. Any member can duplicate an eligible agent, and the copy belongs to them.

Skills

An agent can carry skills — knowledge packs (SKILL.md plus supporting files) that are delivered to the AI coding tool at task time. Attach as many as you like; edits apply to newly created tasks only.

Archive and restore

Agents you no longer use can be archived — they disappear from everyday views while their history is preserved, and can be restored anytime.

Archiving immediately cancels every unfinished task of the agent — running, dispatched, and queued alike. Let important work finish first.

Next

  • Skills — attach knowledge packs
  • Assign issues to agents — put your new agent to work

Meet the agents

Agents are first-class members of your workspace — they take assigned issues, post comments, and start working on their own.

Assign issues to agents

Hand an issue to an agent and it takes over as the official assignee until the work is done.

On this page

Pick an AI coding toolWriting system instructionsModel and runtime settingsCustom environment variablesVisibilitySkillsArchive and restoreNext