Case Study

The Forge

Three AI agents run our entire operation — strategy, engineering, build ops, deployment, content production. They coordinate over Discord, spawn parallel workers, and ship real software without human intervention. This is how it works.

Stack Claude Opus 4.6 + Sonnet 4.6 Platform OpenClaw Uptime 24/7 Autonomous

One human. Multiple products. Not enough hours.

Digital Forge Studios ships multiple software products simultaneously — a cross-agent memory system, an AI-native newsletter, a card game, and internal tooling. Each needs daily attention: code reviews, deployments, content production, infrastructure monitoring, email triage.

A single founder can't context-switch across all of these and still make forward progress. The traditional answer is "hire a team." Ours was different: build agents that operate like a team, with real specializations, real coordination, and real accountability.

Three orchestrators. Parallel workers. One coordination layer.

🪽

Icarus

Strategy · Deployment · Content

  • Product decisions and prioritization
  • WHM/cPanel hosting management
  • Newsletter production pipeline
  • Code review and synthesis
  • Remote deployments (Azure, dforge)
💬

Discord

Coordination Layer

  • Per-project channels (#sulcus, #dforge, etc.)
  • Cross-agent discussion before shared work
  • Automated cron announcements
  • Human oversight and intervention
  • Sync reports and status updates
🔨

Daedalus

Engineering · Infrastructure

  • Rust core development (Sulcus server)
  • CI/CD pipeline management
  • Local builds and testing
  • Godot game engine work
  • Infrastructure code
🧵

Ariadne

Build Engineering · VPS Operations

  • Rust builds and compilation on Forge VPS
  • Container management and Docker operations
  • Minerva zkESG platform development
  • Server-side project operations
  • Direct infrastructure access (Linux/Debian)
↓ spawn workers ↓

Sonnet Sub-agents

Parallel Execution Layer

Each orchestrator spawns Claude Sonnet 4.6 workers for scoped tasks — research, implementation, refactoring, testing. Workers run independently, return structured results, and are killed on completion. Up to 6 concurrent workers per orchestrator.

Territory Rules

Each agent owns specific projects and responsibilities. Icarus handles deployments and remote infrastructure. Daedalus handles Rust core and local builds. Neither touches the other's territory without coordination.

Coordination Protocol

Before any shared work, agents discuss in the relevant project channel and agree on scope. Clear task boundaries prevent overlap. The agent who started first has priority — the other adapts.

Automated Sync

A cron job runs every hour to check for delta between agents — new commits, divergence, territory violations. Zero-noise protocol: only reports when something actually changed.

Heartbeat Loop

Each cycle verifies system health — embedded database, memory operations, sync worker status. Structured status lines with timestamps, actions, and results create an auditable operations log.

What happens while you sleep.

02:00

Gmail Cleanup

Sonnet agent triages the inbox — deletes spam and promotions, archives legal threads, applies labels, flags ambiguous items for manual review. Delivers a morning summary to Discord.

Every hour

Icarus-Daedalus Sync

Delta check across all shared repos. Compares current HEADs to last known state. Flags new commits, divergence, or territory violations. If nothing changed: SYNC_NOOP — zero noise.

Every 30 min

Heartbeat Verification

Checks embedded PostgreSQL, Sulcus plugin health, memory node count, sync worker status, and server request throughput. Structured log lines feed an auditable operations history.

Daily

Newsletter Production

Full pipeline: research sub-agent finds top AI stories → writer sub-agent drafts the issue → build script generates static HTML → FTP deploy to insider.dforge.ca → announce in Discord. Start to finish, zero human intervention.

On-demand

Build, Deploy, Ship

Human says "deploy sulcus-web" or "add Crazy Shapes to the site" in Discord. Agent reads the request, builds context, executes the deployment, verifies live, and confirms. Average deployment: under 5 minutes from request to live.

Production metrics from the first week.

66+ Operations logged Structured status lines since Mar 9
28 Verified passes Zero failures in heartbeat checks
6 Newsletter issues Researched, written, deployed autonomously
6 Live products Booker · Minerva · Sulcus · Insider · Shapes · Forge
3 Cron pipelines Newsletter · Sync · Email triage
0 Human wake-ups Agents handle errors autonomously
"The agents don't need me for daily operations anymore. I intervene for product decisions and strategic pivots. Everything else — deployment, monitoring, content, coordination — runs autonomously."
— Dooley, Digital Forge Studios

Lessons from running a multi-agent operation.

Territory rules prevent chaos

Without explicit ownership boundaries, two agents will inevitably modify the same file at the same time. We define who owns what in a shared coordination document and enforce it through cron-based territory violation checks.

Zero-noise monitoring matters

Early versions of the sync check announced every run — even when nothing changed. This trained us to ignore it. Switching to delta-only reporting (SYNC_NOOP for no change) made every announcement signal, not noise.

Structured logs are the memory

Every operation writes a structured status line: timestamp, project, deliverable, action, result, next step. This creates an auditable trail and lets any agent reconstruct context from cold start.

Opus orchestrates, Sonnet executes

Claude Opus handles decisions, synthesis, and code review. Sonnet sub-agents handle the grunt work — research, implementation, testing. This mirrors human team structures: senior architects direct, junior developers implement.

Discord is the right coordination layer

It provides per-project channels, threading, bot integration, and human oversight in one place. Agents post status updates, discuss before shared work, and receive direct commands — all in the same interface humans use daily.

Fail partially, not totally

When a sub-agent fails, the orchestrator continues with partial outputs and reruns only the failed slice. This prevents one bad API call from blocking an entire pipeline. Recovery is built in, not bolted on.

Built on.

OpenClaw Agent runtime & orchestration
Claude Opus 4.6 Orchestrator model (via Azure Foundry)
Claude Sonnet 4.6 Worker model (parallel execution)
Discord Coordination & human interface
Azure Foundry Model hosting & inference
Sulcus Cross-agent persistent memory
Azure Container Apps Cloud deployment target
cPanel / WHM Static site hosting

The agents are running right now.

While you read this, Icarus, Daedalus, and Ariadne are coordinating on the next deployment, monitoring system health, and planning tomorrow's newsletter. This page was deployed by an agent.