Skip to main content

Orchestrator

Multi-agent task execution with cache coordination.

Overview

Orchestrator supervises agents that claim and execute tasks from a shared cache.

Architecture

Cache keys:

Features

  • Tasks: LLM generates DAG from instructions
  • Locking: Time-limited task leases
  • States: BLOCKED → READY → RUNNING → DONE/FAILED
  • Control: Heartbeat files accept PAUSE, RESUME, DROP-TASK, etc.
  • Spawning: Each agent gets instruction file, heartbeat, goal ID

Files

Configuration

systemg.yaml:
The orchestrator example disables systemg-managed service output capture because agent workloads can be noisy. Use your regular logging pipeline for production agent output, and reserve sysg logs for services that opt into logs.sink: file. Summary status mode avoids expensive process tree discovery across many dynamic children while status and inspect still read current persisted state. PORKI_LLM_MAX_CONCURRENCY raises the number of global LLM leases available across agents, and PORKI_LLM_CONCURRENCY_WAIT_SECONDS controls how long agents wait for a lease before timing out. instructions/INSTRUCTIONS.md:

DAG Format

Usage

Validation

Use sysg status and Redis keys to validate runtime behavior. Package-level runtime tests execute in the porki repository.