Autonomous AI Agent Orchestration

Problem

Managing 28 software projects across a server fleet means constant monitoring and context-switching: services crash, builds fail, deployments need attention, trading bots need P&L summaries, content pipelines stall, and new business opportunities appear on Upwork that need rapid proposals. No single person can monitor all of that in real-time while also doing productive work. The goal was to build an AI system that autonomously manages the entire operation — and that you can talk to from your phone.

Solution

Built a conversational AI orchestrator that manages 28 projects across 6 categories (network engineering, AI automation, trading systems, content pipelines, data engineering, and business operations) through an SMS interface. Text it a question from anywhere — a job site, a coffee shop, the grocery store — and get answers in seconds. It monitors project health, delegates complex work to dedicated child sessions, sends daily morning digests, generates automated forex P&L reports, scans Upwork for relevant jobs, runs weekly research agents, and publishes blog content through a fully automated pipeline.

Architecture

SMS Interface
    │
    ├── Command Router ──→ Action Commands (start/stop/status/etc.)
    │
    ├── AI Router ───────→ Natural Language Processing
    │                         │
    │                    Context Assembly (28 projects)
    │                    Multi-source awareness
    │
    ├── Child Sessions ──→ Dedicated workers per project
    │                         │
    │                    Delegated complex tasks
    │                    Isolated execution environments
    │                    Auto-recovery on failure
    │
    └── Scheduled Systems
            │
            ├── Morning Digest ──────→ Daily project health summary
            ├── Forex P&L Reports ───→ Automated trading summaries
            ├── Upwork Scanner ──────→ Job relevance scoring
            ├── Research Agent ──────→ Weekly industry news scan
            └── Auto-Blog Pipeline ──→ Research → Draft → Review → Publish

The orchestrator maintains awareness of all 28 projects through continuous health scanning. Projects communicate status through signal files — requesting input, reporting completion, or flagging errors. The orchestrator reads these signals, takes action, and notifies via SMS when human attention is needed.

Key Decisions

SMS as the primary interface. Not a web dashboard, not a Slack bot — SMS. It works from any phone, anywhere, with no app to install, no VPN to connect, no browser to open. When a trading bot hits its daily loss limit at 2 AM or a service crashes while you’re at a client site, you get a text. Reply to it and the orchestrator takes action. The latency between “something happened” and “I’m handling it” drops from hours to seconds.

Child session delegation. Complex tasks — debugging a failed build, writing a proposal, analyzing a dataset — get delegated to dedicated child sessions. Each child runs in isolation with its own context, focused on a single project. The orchestrator spins them up, monitors progress, captures results, and cleans up. This is how a single orchestrator scales to 28 projects without drowning in context.

Auto-recovery and health monitoring. Services crash. Builds fail. Network connections drop. Rather than alerting a human for every transient failure, the orchestrator attempts automatic recovery first — restarting services, retrying failed operations, clearing stale state. Humans only get notified when auto-recovery fails or when the situation requires a judgment call.

Tiered notification system. To prevent notification fatigue, alerts are classified by urgency. Critical events (service down, trading limit hit) bypass quiet hours. Routine summaries (daily digest, P&L reports) batch into scheduled deliveries. Debug-level events log silently. A daily SMS budget prevents runaway notifications from flooding your phone.

Upwork job scanning with relevance scoring. The orchestrator monitors Upwork job feeds, scores listings against a profile of ideal engagements (network automation, AI integration, Python/FastAPI), and surfaces high-relevance opportunities via SMS. This turns passive job browsing into an active notification system — relevant work finds you instead of you hunting for it.

Auto-blog pipeline. A fully automated content pipeline that runs weekly: research agents scan industry news and technical developments, draft articles based on findings, run editorial review passes, and publish to the blog. No human touches the content unless quality gates flag an issue for review.

Results

28 projects managed across 6 categories from a single SMS interface
Daily morning digests summarizing project health, pending tasks, and alerts
Automated forex P&L reports delivered on schedule
Child session delegation for complex tasks with isolated execution
Auto-recovery handling transient failures before alerting humans
Upwork job scanning with relevance scoring and SMS notifications
Weekly research agent scanning industry news and developments
Auto-blog pipeline: research → draft → review → publish
Tiered notification system with daily SMS budget and quiet hours
Configurable autonomy levels: observe, cautious, moderate, full

How This Scales

Multi-user access — Extend the SMS interface to support multiple operators with per-user permissions and project scoping.
Analytics dashboard — Web UI showing session history, project health trends, notification logs, and autonomy level progression over time.
Client deployments — Package the orchestrator for teams managing fleets of AI agents, with per-organization configuration and notification channels.
Automatic trust promotion — Track session success rates, error frequency, and task completion to automatically increase autonomy levels when the system proves reliable.
Webhook integrations — Beyond SMS, support Slack, Discord, and email notification channels for team environments.

Tech Stack

Orchestrator: Node.js
AI Engine: Claude CLI (child process delegation with concurrency limits)
SMS: Native macOS messaging (no third-party dependencies)
State: SQLite, JSON-based state persistence
Scheduling: Automated service management with health monitoring
Infrastructure: Mac Mini (always-on server)