Skip to main content
Before deploying to production, your release gate agent collects evidence, summarizes risk via an LLM tool call, and may trigger a CI/CD pipeline. This recipe models that as a pipeline session with explicit high-risk metadata. Source: production-release-gate-loop.ts · production_release_gate_loop.py

​
Scenario

A release gate orchestrator validates rollout readiness. It calls an LLM tool to summarize risk, then triggers a CI/CD pipeline via MCP. Both actions need production metadata for guardrails.

​
Step 1 — Pipeline session

await apie.withSession(
  {
    kind: "pipeline",
    inputSummary: "Validate production rollout readiness",
    metadata: { workflow: "release_gate" },
  },
  async (session) => { /* orchestrator run inside */ },
);

​
Step 2 — Orchestrator run with step metadata

await apie.withRun(
  {
    sessionId: session.id,
    stepKey: "orchestrator",
    stepName: "Release gate orchestrator",
    stepIndex: 0,
    inputSummary: "Collect rollout evidence and risky actions",
  },
  async (run) => { /* tool calls inside */ },
);

​
Step 3 — LLM tool call (medium risk)

await withOpenAIToolCall(apie, {
  runId: run.id,
  toolName: "summarize_release_risk",
  arguments: { service: "api", environment: "production" },
  resourceType: "work_item",
  riskLevel: "medium",
}, async () => ({ summary: "No blocker found in latest incident feed." }));

​
Step 4 — MCP CI/CD trigger (high risk)

await withMcpToolCall(apie, {
  runId: run.id,
  sessionId: session.id,
  server: "internal-cicd",
  tool: "trigger_pipeline",
  actionType: "execute",
  resourceType: "pipeline_run",
  environment: "production",
  riskLevel: "high",
  resourceTarget: "payments-service",
}, async () => ({ accepted: true, runId: "pipe_123" }));

​
What you’ll see

A pipeline session replay with orchestrator run, LLM tool call (medium risk), and MCP pipeline trigger (high risk). In Monitor mode, guard evaluations show what would be blocked in Enforce mode.

​
Escalate to enforcement

  1. Declare capabilities for trigger_pipeline and summarize_release_risk
  2. Enable guardrail templates for prod-deploys
  3. Switch to Enforce mode

​
Next steps

Multi-agent handoff

Orchestrator delegates to worker.

MCP instrumented client

In-process MCP setup.