AI Agents

Orchestra AI agents are autonomous teammates deployed within a workspace. Each agent has a defined role, authorized tools, custom instructions, and optionally a model override. Agents execute tasks asynchronously via the background Worker service.

🤖 AI Teammates 10 agents
+ New Agent
● Busy CLI Agent
đŸŽģ Maestro
Coding Agent ¡ github-copilot
14 tools ¡ 2 skills ¡ 1 active job
● Idle API Agent
đŸŽĩ Aria
Code Review Agent ¡ gpt-4o
8 tools ¡ 1 skill ¡ 0 active jobs
● Idle CLI Agent
🔍 Orpheus
Agentic Search Agent ¡ claude-haiku-4.5
6 tools ¡ 3 skills ¡ 0 active jobs
● Idle API Agent
📋 Thalia
Code Review Agent ¡ gpt-4o-mini
5 tools ¡ 0 skills ¡ 0 active jobs
The AI Teammates page — deploy and manage autonomous agents

Built-in Agent Templates

Orchestra ships three production-ready templates to get you started immediately:

Built-in
🔍 Code Review Agent

Automated PR/MR analysis via GitHub or GitLab. Operates in read-only mode — reviews diffs, comments on quality, style, and security issues.

  • Read-only repository access
  • PR diff analysis
  • Automated comment generation
Built-in
🔎 Agentic Search Agent

Deep codebase exploration and summarization. Uses read-only CLI access to traverse repositories and answer complex technical questions.

  • Read-only CLI access
  • Codebase exploration
  • Technical summarization
Built-in
đŸ’ģ Coding Agent

End-to-end implementation with full read/write access. Handles feature development, bug fixes, and refactoring tasks from ticket to commit.

  • Full read/write CLI access
  • File creation and editing
  • End-to-end implementation
💡

Template versioning: Built-in agents track a TemplateVersion at creation time. When the template is updated in a new Orchestra release, you can re-deploy to get the latest instructions while keeping your existing tool authorizations.

â„šī¸

Not enough? In addition to built-in agents, Orchestra offers Starter Templates — customizable blueprints for Product Planning, Development, QA, and Business Analysis. All fields are fully editable and templates use your workspace's AI provider by default.

Creating an Agent

Navigate to Agents in the workspace sidebar and click Deploy Agent. You'll be presented with two paths:

A
Use a Built-in Template

Browse the template catalogue. Templates come pre-configured with appropriate tools and instructions. Some fields are locked to preserve template integrity.

B
Create from Scratch

Define a custom agent with a name, role, capabilities, custom instructions, and manually select tool authorizations from the full tool library.

Agent form fields

FieldDescription
NameDisplay name for the agent (e.g., "Senior Code Reviewer")
RoleShort role description used in system prompts (e.g., "Senior Software Engineer")
CapabilitiesComma-separated tag list (e.g., code-review, typescript). Used for filtering.
Custom InstructionsFree-form system instructions for the agent's behavior (mutually exclusive with Project Principles)
Project PrinciplesCode review-specific instructions — code standards, review focus areas (mutually exclusive with Custom Instructions)
Model OverrideOptional: override the workspace default model for this specific agent

Tool Authorization

Orchestra enforces granular tool authorization at the ToolAction level. Each tool (e.g., "GitHub Integration") contains multiple actions (e.g., "Create Issue", "List Pull Requests"). Agents only have access to the specific actions you authorize.

🔧 Native Tools

Built-in integrations: GitHub, Jira, GitLab. Actions are predefined and use reflection-based invocation in the backend.

⚡ MCP Tools

Tools discovered from connected MCP servers. Appear alongside native tools in the tool picker with a distinct MCP badge.

Danger Level Classification

Every tool action carries a danger level to prevent unintended destructive operations:

Safe

Read-only operations — list, get, search. No side effects.

Moderate

State-modifying but reversible — create issue, post comment, update status.

Destructive

Irreversible operations — delete branch, force push, drop data.

Custom Instructions vs Project Principles

â„šī¸

Mutually exclusive: An agent uses either Custom Instructions or Project Principles — never both. The correct field is automatically selected based on whether any code review tool actions are authorized.

FieldUse WhenExample Content
Custom Instructions General-purpose agents (coding, search, triage) Coding style preferences, technology constraints, response format requirements
Project Principles Code review agents SOLID enforcement, naming conventions, security rules, test coverage requirements

Model Selection

By default, agents use the workspace's configured AI provider model. You can override this per-agent to use a different model for cost/quality trade-offs:

  • Azure OpenAI: Select from deployed models in your Azure resource
  • Ollama: Select from models discovered via ollama list
  • GitHub Copilot: Configure reasoning effort level (low / medium / high) per CLI integration