Orchestrator agents that call sub-agents can use mocked fixtures to test the full LLM reasoning loop without requiring live sub-agents. Create fixtures with a mocks field that maps custom tool names to deterministic responses:
The workflow is the same — orch test automatically detects mocked fixtures and runs the full agent loop with mock responses injected for custom tool calls. No additional CI configuration needed.
Mocked orchestration tests verify that the LLM correctly selects tools, processes their responses, and produces the expected output — all without network calls to deployed sub-agents. This makes them fast and reliable in CI.
Required Secrets
Add these to your GitHub repo under Settings > Secrets and variables > Actions:
Secret
Description
ORCHAGENT_API_KEY
Your orchagent API key (from orch login or dashboard)
ANTHROPIC_API_KEY
LLM key for running fixture tests and mocked orchestration tests
Deploying Multiple Agents
For monorepos with multiple agents, use --all to publish everything in dependency order:
This scans subdirectories for orchagent.json/SKILL.md files, builds a dependency graph, and publishes leaf-first. If any agent fails, the batch stops.Alternative: For more control, use a matrix strategy:
Matrix strategies run agents in parallel — ensure leaf agents are published before orchestrators by setting max-parallel: 1 or splitting into separate jobs with dependencies.
Version Pinning
By default, publishing auto-increments the version (v1, v2, v3…). Schedules and webhook triggers auto-update to the latest version unless pinned.To pin a schedule to a specific version: