orchestrate

name	orchestrate
description	Spawn and manage autonomous agent fleets using initialPrompt + /schedule + /loop + --channels + auto mode. Define agent configurations, standing orders, and schedules. The glue between all orchestration primitives.
updated	"2026-03-25T00:00:00.000Z"
triggers	["orchestrate","spawn fleet","agent fleet","schedule agents","autonomous agents","standing orders","fleet mode","spawn specialists","recurring agents","agent schedule"]

Orchestrate — Agent Fleet Management

Spawn specialist agents with standing orders, schedules, and autonomous execution. Combines Claude Code primitives (initialPrompt, /schedule, /loop, --channels, auto mode) into managed workflows.

The Primitives

Primitive	What It Does	Claude Code Feature
initialPrompt	Agent starts itself — no human types first message	Agent frontmatter
/schedule	Agent runs on a cron schedule	Remote triggers
/loop	Agent runs on a recurring interval	Polling loop
--channels	Agent receives events from Telegram/Discord/webhooks	Channel MCP servers
--permission-mode auto	AI classifier approves safe ops, escalates risky ones	Permission mode
Skills	Agent loads domain-specific instructions and tools	Skill system

Fleet Patterns

Pattern 1: Scheduled Specialist

Agent runs on a schedule with standing orders. No human involvement.

Agent: security-auditor
  initialPrompt: "Run codebase-healthcheck. Audit top 5 hotspots for OWASP Top 10."
  schedule: "0 9 * * 1"  (Every Monday 9am)
  permissionMode: auto
  skills: codebase-healthcheck, session-memory

Pattern 2: Monitoring Loop

Agent checks something repeatedly and acts on findings.

Agent: deployment-monitor
  initialPrompt: "Check all Vercel deployments. Report failures. Store findings in session-memory."
  loop: 30m
  permissionMode: auto
  skills: vercel-deployment, session-memory

Pattern 3: Event-Driven via Channels

Agent sits idle until an external event arrives (Telegram message, webhook, N8N trigger).

Agent: alert-responder
  initialPrompt: "You monitor this channel for alerts. When one arrives, triage it, check the affected system, and respond with diagnosis and recommended action."
  channels: plugin:telegram
  permissionMode: auto
  skills: session-memory

Pattern 4: Orchestrator Spawning Specialists

Parent Claude spawns multiple agents, each with their own role.

# Parent orchestrator spawns the fleet:

# Security audit every Monday
/schedule create --name "security-audit" --cron "0 9 * * 1" --agent security

# QA check after every deploy (triggered by N8N webhook)
# N8N watches Vercel deploy webhook → sends to channel → QA agent picks up

# Billing reconciliation on the 1st
/schedule create --name "billing-reconcile" --cron "0 10 1 * *" --agent invoice-automation

# Performance monitoring every 6 hours
/loop 6h "Spawn performance agent against production URL"

How to Set Up a Fleet

Step 1: Choose agents and standing orders

Decide which agents run, what their initialPrompt is, and what schedule they need. Each agent should have a clear, specific first instruction — not "do stuff" but "run X tool, check Y, report Z."

Step 2: Configure auto mode

Ensure --permission-mode auto is configured in settings with appropriate allow/deny rules. Without this, agents get blocked by permission prompts.

Step 3: Set schedules or loops

# Cron-based (runs at specific times)
/schedule create --name "weekly-security" --cron "0 9 * * 1" --agent security

# Interval-based (runs every N minutes/hours)
/loop 30m /some-check-command

Step 4: Wire external triggers (optional)

For event-driven agents, connect N8N webhooks or --channels:

• N8N receives external event (NinjaOne alert, client email, Stripe webhook)
• N8N sends to Telegram group or direct webhook
• Agent with --channels picks up the event and acts

Step 5: Monitor via session-memory

All agents should use session-memory to store their findings. The orchestrator (or you) can search across all agent memories:

python3 ~/.claude/tools/session-memory/memory.py search --query "security vulnerabilities found" --limit 10
python3 ~/.claude/tools/session-memory/memory.py search --query "deployment failures" --limit 10

Agent Configuration Reference

Agents declare their standing orders in frontmatter:

---
name: my-agent
description: What this agent does
tools: Read, Grep, Bash, Sequential
initialPrompt: "Read CLAUDE.md. Run codebase-healthcheck. Report top 5 findings."
permissionMode: auto
---

Safety

• Auto mode guardrails: Each agent operates under the same allow/deny rules
• Session-memory audit trail: Every agent stores what it found and decided
• Safety limits: Non-stop mode caps apply (iterations, cost, duration)
• Channel escalation: --channels routes approval prompts to your phone
• No self-modification: Auto mode blocks agents from changing their own config

Integration Points

• Chains to: All agent definitions (via initialPrompt), non-stop (autonomous execution), session-memory (audit trail)
• Chains from: N8N (external triggers), --channels (Telegram/Discord), /schedule (cron), /loop (intervals)
• Server stack (example): workflow engine (e.g. N8N) + chat bridge (e.g. Telegram bot or Slack app) + agent fleet manager (your orchestrator of choice). Substitute your own components — the orchestrate skill is agnostic to the specific tools at each layer.