Glink

Multi-Agent Workflow Orchestration. One Bus. Zero Friction.

Glink is a lightweight orchestration engine that turns your AI agents into a collaborative assembly line. Define a workflow in YAML, and Glink routes each step to the right agent — passing context, handling failures, and logging every heartbeat onto a shared event bus.

No database. No message queue. No external dependencies.

---

Architecture

┌──────────────────────────────────────────────────────────┐
│                     Glink Daemon                          │
│                                                           │
│  ┌─────────────────────────────────────────────────────┐  │
│  │                    Main Bus                          │  │
│  │              Append-only JSONL Timeline              │  │
│  └──────┬──────────┬──────────┬──────────┬─────────────┘  │
│         │          │          │          │                │
│    ┌────▼───┐ ┌───▼────┐ ┌───▼────┐ ┌──▼──────┐         │
│    │Agent-1 │ │Agent-2 │ │Agent-3 │ │...      │         │
│    │:8420   │ │:8431   │ │:8432   │ │         │         │
│    └────────┘ └────────┘ └────────┘ └─────────┘         │
│         │          │          │          │                │
│         └──────────┴──────────┴──────────┘                │
│                    Your AI Fleet                          │
└──────────────────────────────────────────────────────────┘

---

Quick Start

# Define your agents in main.py or daemon/core.py
# Default port mapping (customize freely):
#   agent-1 :8420 (generalist)
#   agent-2 :8431 (backend)
#   agent-3 :8432 (frontend/UI)
#   agent-4 :8434 (data)
#   agent-5 :8435 (testing)

# Run a workflow (auto-resumes from last checkpoint)
python3 glink-daemon.py my-workflow

# Force restart from step 1
python3 glink-daemon.py my-workflow --force

# Jump to a specific step
python3 glink-daemon.py my-workflow --step 4

# Serve-only mode (API daemon without running workflow)
python3 glink-daemon.py --serve

# Open dashboard
open http://127.0.0.1:8426/commander.html

---

Features

Feature	Description
YAML Workflows	Define steps, agents, dependencies, and fallbacks in one file
Main Bus	JSONL blackboard — append-only, agent-agnostic, replayable
Smart Routing	Primary agent down? Auto-fallback to the next in line
Checkpoint Resume	Crash mid-workflow? Restart picks up where it left off
Dependency Graph	Steps can depends_on each other; Glink handles ordering
Retry Loop	Auto-retry failed steps (configurable, default 2×)
HTTP API + SSE	Live status, agent health, and event stream on :8426
Self-Healing	Daemon auto-restarts on crash, PID-based watchdog
Webhook Alerts	Push notifications to any HTTP endpoint
Zero External Deps	Pure Python 3.10+, standard library. No pip install.

---

Define a Workflow

name: my-pipeline
version: 1.0.0
description: "A simple 3-step demo"

global_context: |
  You are part of a multi-agent orchestration pipeline.
  Build upon the output of previous steps.

steps:
  - id: step-1
    executor: agent-1
    title: "Generate content"
    output_file: projects/demo/step1.txt
    task: |
      Create a summary of what makes a good multi-agent workflow.

  - id: step-2
    executor: agent-2
    title: "Enhance"
    input_file: projects/demo/step1.txt
    output_file: projects/demo/step2.md
    task: |
      Read and enhance the step-1 output. Add code examples.

  - id: step-3
    executor: agent-3
    title: "Verify"
    input_file: projects/demo/step2.md
    output_file: projects/demo/VERIFIED.md
    task: |
      Verify the enhanced document is complete.
      Append a verification seal if all checks pass.

---

API Reference

All endpoints served on the configured port (:8426 by default).

Method	Endpoint	Description
GET	/health	Liveness check
GET	/status	Full project status + step-by-step progress
GET	/status/agents	Which agents are online
GET	/status/events?n=20	Last N bus events
GET	/intel/step	Detailed intelligence per step stage
GET	/intel/agents	Agent-specific metrics
GET	/intel/timeline	Step timeline visualization data
GET	/events/stream	SSE real-time event stream
POST	/restart	Resume from last checkpoint
POST	/restart?force	Force restart from step 1
POST	/restart?step=N	Jump to specific step

---

Configuration

See glink-config.yaml:

project:
  default: hello-world

scheduling:
  max_retries: 2
  poll_interval: 3
  poll_max_wait: 180
  max_concurrent_steps: 1

reporting:
  channels:
    - type: console
      label: "Glink"
    # - type: webhook
    #   url: "https://hooks.example.com/..."
    #   label: "Slack"

server:
  host: "127.0.0.1"
  port: 8426

security:
  startup_timeout: 10

Environment variables:

• GLINK_DEFAULT_PROJECT — override default project name
• GLINK_PORT — override API server port
• GLINK_REPORTER — set to webhook, console, or silent
• GLINK_ALERT_WEBHOOK — webhook URL for alerts

---

Real-World Usage

Glink was used to orchestrate a 10-step game development pipeline across 5 agents:

• Step 1-4: 3D scene, physics, textures, UI
• Step 5-6: Game systems (save/load, scoring)
• Step 7-8: Quality verification
• Result: Single playable HTML file, 97 KB / 2,751 lines

All built by agent collaboration — zero lines of human-written code.

---

Project Structure

glink/
├── glink-daemon.py         # CLI entry point
├── glink-config.yaml        # Configuration
├── daemon/
│   ├── core.py              # Workflow orchestration engine
│   ├── api.py               # HTTP API server (17 endpoints)
│   ├── checks.py            # PID management & auto-recovery
│   ├── config.py            # Config loader
│   └── log.py               # Reporter initialization
├── bus/
│   ├── main_bus.py          # JSONL event bus
│   └── agent_client.py      # Agent HTTP client
├── reporter/
│   └── reporter.py          # Notification session (webhook/console)
├── dashboard/
│   ├── commander.html       # C2 dashboard (realtime)
│   └── index.html           # Legacy dashboard
├── workflows/               # Your YAML workflow definitions
└── projects/                # Step outputs by project

---

License

MIT — free for any use, open or commercial.