Hivemind

The Problem

AI coding assistants are powerful, but deploying them at scale is a mess. Every task needs a clean environment, the right context, and proper isolation. Running AI agents manually — one terminal at a time — doesn’t scale.

The Solution

Hivemind is a full-stack platform built in F# (SAFE stack) that manages the entire lifecycle of AI coding agents:

• Container Orchestration: Each coding task runs inside an isolated Docker container with the right repo, branch, and dependencies pre-loaded.
• Agent Management: Supports multiple AI backends — Claude Code, Codex CLI, and others — with a unified interface for spawning, monitoring, and collecting results.
• Pipeline Automation: Multi-phase workflows (plan → implement → review → fix → PR) run automatically, with human checkpoints where they matter.
• Fleet Operations: Run dozens of coding agents in parallel across different repos, branches, and tasks.

Architecture

┌─────────────┐     ┌──────────────┐     ┌────────────────┐
│  Hivemind   │────▶│  Agent       │────▶│  Docker        │
│  API Server │     │  Orchestrator│     │  Containers    │
└─────────────┘     └──────────────┘     └────────────────┘
       │                    │                     │
       ▼                    ▼                     ▼
┌─────────────┐     ┌──────────────┐     ┌────────────────┐
│  Web UI     │     │  Git Manager │     │  AI Agents     │
│  (Fable)    │     │  (Gitea)     │     │  (Claude/Codex)│
└─────────────┘     └──────────────┘     └────────────────┘

Key Decisions

• F# / SAFE Stack: Strong typing catches orchestration bugs at compile time. The functional-first approach makes complex state machines manageable.
• Docker isolation: Every agent runs in its own container. No cross-contamination between tasks. Clean environments every time.
• Git-native: Everything flows through git. Plans are committed to specs/, implementations happen on feature branches, reviews happen on diffs.

Impact

Hivemind has become the backbone of my development workflow. What used to take hours of manual context-switching — spinning up environments, feeding context to AI tools, reviewing output — now happens automatically while I focus on architecture and code review.

Status

Active development. Currently used as the primary development infrastructure for multiple projects.