joelhooks/swarm-tools

What it solves

Swarm solves the problem of "context death" and coordination overhead when using AI agents for large, complex tasks. It prevents agents from conflicting with one another when editing files and ensures that progress, task tracking, and learned patterns are preserved across different sessions.

How it works

Swarm uses a coordinator to decompose a large task into smaller, parallelizable subtasks. It then employs several core systems to manage the workflow:

• Hive: A git-backed task tracker that stores tasks in a .hive/ directory, allowing progress to survive session restarts and sync across environments.
• Hivemind: A semantic memory system using Ollama embeddings to store and retrieve learnings and patterns from previous tasks.
• Swarm Mail: An actor-model coordination system that handles agent-to-agent messaging and implements file reservations (locks) to prevent multiple agents from editing the same file simultaneously.
• Event Sourcing: All state changes are recorded in an append-only event log using libSQL (embedded SQLite), ensuring a durable record of every action.

Who it’s for

Developers using AI coding assistants like OpenCode or Claude Code who need to scale from single-prompt tasks to complex, multi-agent workflows that require persistence and coordination.

Highlights

• Parallel Execution: Spawns multiple worker agents to handle subtasks simultaneously.
• Persistent Memory: Uses a semantic memory system to record successful patterns and identify anti-patterns.
• ** Poirot-style Coordination:** Implements file locking to avoid merge conflicts between agents.
• Local-First Architecture: Runs entirely locally with no external servers, utilizing Bun, libSQL, and Ollama.